Your AI context doesn't have to stay trapped inside one app. With MCP (Model Context Protocol), you can bring your entire ambientChat world into Claude Desktop, brand apps, or any compatible AI client — and it works exactly like having a very well-informed research assistant sitting at your desk.

All tools respect your authentication. The AI can only access your data, using your credentials. Nobody else's stuff, ever.


list_items with search and search_items also return owned Inventory folders in matching_folders, including empty/nested namesakes. Show each full folder_path_label; open the returned view ID as an Inventory folder. Folder matches are separate from item counts and caps. list_items returns discovery on page one only; includeFolders:false omits it for an item-only read. If folder_search_failed is true, item results remain valid but no claim about absent folders is justified.

#Inventory Tools

Manage your physical items — everything you've scanned, photographed, or manually added.

Tool What It Does
list_items List all items, optionally filtered by category, folder, or search query. reaction_first: true puts every liked item in scope first on page 1 (the Favourites strip) and never repeats one on a later page. Finished items are left out; page 1 says how many with finished_count, and include_finished: true lists them too (each with status: "finished")
get_item Get full details of a specific item — name, brand, price, expiry, location, notes, nutritional data
search_items Semantic search — finds items by meaning, not just keywords ("things in my garage", "snacks for the party")
list_expiring_items Find items expiring within a number of days (default 14), soonest first. Set include_expired to also see items already past their date; limit sets the page size (default 200) and truncated flags when more match
get_items_at_scene List items last seen at a specific physical location — identified by beacon, GLN, or scene ID
get_scan_session_items List items created together in the same scan session (scan siblings)
update_item Edit any item field — name, brand, category, price, expiry date, notes
finish_item Mark an item Finished: used up, given away, sold, broke or lost (reason, optional). It leaves every "what I've got" answer but keeps its photos, documents, reaction and history. Items carry status, finished_reason, finished_at and, for a sale, finished_sale_price. Pass finished_sale_price with reason: "sold" (or again later, on an item already sold). The reply carries the same follow-ups the apps' finish sheet shows: warranty_documents for a broken item and a follow_up line (note the sale price, name the warranty, offer to find a lost thing first, offer the shopping list for something used up or broken — never for something they disliked). Pass add_to_list: true to also put it on their shopping list (the list they last added to, a new "Shopping list", or the one you name with list_id), linked to the item: the reply's list_row names the line and the list — tell them which, and already_on_list is true when an unticked line for it was already there, so nothing was added twice. Undo with undo_item_change on its event_id. "We drank the Opus One", "used up the olive oil", "sold the bike for $300"
unfinish_item "Got it again" — bring a finished item back into the inventory, the same record. "I bought more of that olive oil", "found the keys after all"
set_reaction Give one of your items or documents a 👍 (reaction: "like") or 👎 ("dislike"), or clear it (null). Items and documents carry reaction and reacted_at. It only records taste — it never moves, finishes or deletes anything. "I loved that Opus One", "that recipe was great", "never buying that coffee again"
get_taste Your taste: everything you gave a 👍 (likes) or 👎 (dislikes) — items and documents, newest first, each with its taste_kind (wine, music, art, food, books, places, other) and, for an item, whether it is finished — plus your interests. A liked item that is finished is one you no longer have: agents recommend it when you are out (a restaurant, a shop, a gift), never for what to cook or use at home; a dislike is never recommended. limit caps each list (default 100); likes_count and dislikes_count say how many there are; interests_unavailable means the interests could not be read just now (the reactions still come back). summary is the short distilled taste summary (text, and source: generated, or edited — your own words), or null when there isn't one yet. Your own reactions only. "What do I like?", "recommend a wine for tonight's restaurant"
set_item_location Set where an item IS — pick one of your zones or type a place. Say "the kayak is in the garage"
update_items_by_filter Bulk-update fields across many items in one call, by category, a search, a folder, "everything in Home", or "everything in any folder". Say "change all items in the kitchen to category Food"
save_web_pointer Keep a scanned web page as a document you can rename and share. Saving the same address twice returns the same document, so your title sticks
read_web_page Read what a page you scanned or saved says RIGHT NOW — a temperature reading, an alert, a date — alongside the title and notes you gave it
promote_item Keep an item the AI Lookup camera created provisionally — this is what Save does
discard_scan_session Throw away an AI Lookup capture nobody kept, with its scan photo. Only ever touches items still awaiting that choice
trash_item Move a single item to trash (soft delete). Documents linked only to it go to trash with it and come back when it is restored
trash_items Bulk move multiple items to trash at once
restore_item Restore an item from trash back to the inventory
regenerate_item_image Generate a fresh AI illustration for an item. Costs credits
verify_item_value Check an item's value against the web: one search, read for up to three prices real pages state, kept on the item as a range with each page, its exact words and the date. Something owned more than about six months is checked for what it would sell for (used, sold or resale prices); if the web only states a new price, the check says that is the replacement cost, not the resale value. Items valued at $250 or more, or about the same in their own currency (€250, £200, ¥40,000…); costs one web search. Never changes the value by itself. "Verify the value of my espresso machine", "back up my TV's value for insurance"
apply_value_check Set an item's Value to the price its last web check found, with that page as its source. Only when you say so; Cost never changes; Undo puts the old Value back. "Use that price"
apply_receipt_price Set an item's Price to its line on the receipt linked to it, when the receipt was offered rather than applied ("Receipt says $X · Use this": you typed the price, or two lines could be the item). Only when you say so; Value never changes; Undo puts the old Price back. "Use the receipt price"
upload_item_image Set the user's own picture (base64) as an item's image — it becomes the thumbnail in folders, chat cards and Radar tour exhibit cards. The photo it replaces (original_image_url) and the AI icon are kept
set_item_face_image Choose which of an item's own pictures is its thumbnail. pin: true makes it the fixed default (face_pinned_url), which an unpinned choice can't move; pin: false clears it
trash_item_image Move one of an item's pictures to Trash. If it was the thumbnail, the next picture takes over. restore_trash with the returned trashId puts it back
find_album_covers Replace AI pictures on albums (vinyl, CD, LP) with their real cover from MusicBrainz / Discogs, and fetch their track listings. Pass a folder (view_id) or itemIds. Albums that already have their cover but no track listing get one. Free; covers and details land over a few minutes. Never replaces a user's photo
revisit_item_match Take a closer look at which record an album is: one paid AI look compares the user's own photo of their copy with the candidate sleeves and recommends one, with a sentence or two of reasoning. Costs about get_dpp → music.match.revisit_credits credits (a cent or so), charged only when the AI actually looked; out of credits it says so and charges nothing. Never switches the record: say the recommendation, and only if the user agrees call set_item_edition with its ref. Offered while the match is open (music.match.can_revisit). "Take a closer look at which pressing my Help! is"
set_item_edition "Which one is yours?" — set which record an album is, after asking the user: ref is one of get_dpp → music.match.candidates[].ref (which album) or editions[].ref (which pressing). The album moves at once; its track listing, year, label and sleeve follow in seconds (music.match.switching until then). Passing the record it is already matched to confirms it (confirmed: true): the doubt clears and nothing moves. The user's choice is final — no automatic re-match changes it. Undo with undo_item_change on the returned event_id. Free. For an album in doubt with fewer than two candidates, pass gather: true instead of ref to start its one free lookup for other editions; get_dpp → music.match.gathering is true while it is on its way (get_dpp itself never starts it). When none of the candidates is theirs, pass hint instead — the catalogue number on their label ("PMC 1255") or the barcode digits: ONE free MusicBrainz search, and a hit is added to music.match.candidates (music.match.hint reads searching, then found with its ref, none or unanswered) — never picked; ask, then pass its ref. "My Help! is the LP, not the EP"
improve_item_photos Clean up items' own photos: crop to the item, turn it upright and evenly fix the light — the camera's own pixels, nothing re-drawn, so the result is fine for resale and insurance. It becomes the item's picture (pinned as its default) and the original stays in its gallery. Costs credits: call it without confirm for the price; at $0.50 or less go ahead with confirm: true, above that tell the user and confirm once they agree. Pass itemIds or a folder (view_id); items with no photo of their own are skipped. mode: "beautify" instead has AI redraw the photo as a studio product shot — display only, never used for resale or insurance; priced at the image rate and on the paid plans (PAID_PLAN otherwise)

Example things you can say:

  • "What's in my pantry?"
  • "Show me everything expiring this week."
  • "Change the category of all my vitamins to Health."
  • "How much is my electronics collection worth?"
  • "What was scanned with the olive oil?" (scan siblings)
  • "What did I leave in the garage?" (items at scene)

#Document Tools

Navigate, search, and manage your digital paperwork — receipts, manuals, photos, and files.

Tool What It Does
list_documents List documents with optional filters for media type, category, and folder. With search, page 1 also includes matching_folders: all matching owned document folders, including empty/nested folders, with full canonical paths. Folder matches are separate from document counts and pagination. reaction_first: true puts liked documents first on page 1
get_document Get full document details — AI summary, linked items, metadata
search_documents Search document titles, summaries, and content, plus owned folder names in matching_folders. Folder results carry IDs and full paths; show them alongside documents and open as view/id. folder_search_failed distinguishes a failed folder read from no matches
search_documents_semantic AI-powered semantic search — finds documents by meaning, returns full content
update_document Update a document's title, description, and/or tags. Improves search and accessibility
upload_document Upload a new document (image or PDF) into a folder for AI processing. Provide the bytes one of three ways: a Firebase Storage URL, a get_document_upload_url storage path, or inline base64 (max 10MB)
get_document_upload_url Get a short-lived signed URL for uploading document bytes directly to storage, then hand the returned path to upload_document. This is how an AI client that holds a file (e.g. one with a code-execution sandbox) gets it into a folder without inlining the bytes
get_scan_status Check the status of an upload/scan job by its jobId (returned by upload_document). Tells you whether it's still processing, complete (with the created document/item), or failed (with the reason) — so you don't have to keep refreshing the folder. When the scan carried an “only capture…” instruction it also returns capture_gate: what was kept, what was skipped and why, whether each reject has since been brought back, and — if the instruction was not applied — not_applied_reason saying whether the photo simply held one thing (nothing to filter) or the filtering step could not run
check_scan_affordability Can the user afford a photo scan right now? The same check the app runs before every scan. When warn is true, a typical photo scan costs more than they have left (expected_credits against credits_remaining) or they are out of credits — say so before they scan and offer an upgrade. When it's false, say nothing about cost
restore_skipped_objects Bring back objects a scan's “only capture…” instruction skipped. Read the list from get_scan_status and pass the positions the user wants, or omit them to restore everything. The object is re-run through the original photo, so what appears is the item the scan would have made — same scan session, same place. Costs credits, about what that object would have cost inside the scan; restoring the same one twice does nothing
move_document Move a document between folders
link_document_to_item Attach a document you already have to one of your inventory items — the manual, warranty, receipt, or spec sheet that belongs with a thing you own. A receipt sets the item's Price to its line on it, unless you typed the price
unlink_document_from_item Detach a document from an item. Removes only the link; both records stay
add_document_pages Add the photos you attached to this message as extra pages of an image document — how a photographed multi-page manual becomes ONE document instead of a pile of photos
combine_documents Gather several photo documents into ONE multi-page document — the fix for a manual uploaded page by page. The first keeps its identity; the rest become its pages and go to Trash
remove_document_page Remove one page from a multi-page document by its number
get_item_documents List the paperwork attached to an item — its manual, warranty, receipt, or the scan photo it came from
trash_document Move a document to trash (soft delete). Cleans up linked inventory items
restore_document Restore a document from trash
regenerate_document_image Generate an AI illustration thumbnail for a document. Costs credits
generate_image Create an image from a text description — optionally grounded in your own photos. Say "draw my armchair in a jungle" and the armchair's actual photo is shown to the generator so the result looks like your armchair, not a generic one; "me as an astronaut" uses your profile photo. The per-image price is checked against your balance before generating, only delivered images are charged, and the result is saved into your context as an image document — with a recipe showing exactly what went into the prompt (and whether each reference was matched from its photo or just described) and a scorecard grading the image against your brief. Say "dual" or "show me both" to run the same brief on Gemini AND OpenAI side by side and pick your favorite. Ask for "widescreen" (3:2) or "phone wallpaper" (9:16) shapes, and say "final quality" for the premium render (paid plans). Prefer knobs over words? Open Image Studio — the ✦ button in chat, or the sidebar on web — for a full composer with model, quality, and shape controls plus live pricing. To build on something you already have, add it as a reference from My Stuff inside the composer
list_image_models The image models generate_image can use and what each costs — the same $×N badge Image Studio shows (how many times more than the cheapest image model), with the exact credits per square and widescreen/tall image, your remaining credits, and whether Final quality is open to you. Free to call. Try: “which image model is cheapest?”, “what would a Sunburst image cost me?” or “which model is best at lettering?”
resolve_dual After a Dual image run, record your verdict: pick a winner (the other image moves to Trash, recoverable) or keep both — no extra charge either way. Works by voice too: "keep the Gemini one", "keep both"

Example things you can say:

  • "Make me an image of my armchair reupholstered in mustard."
  • "Generate a picture of a lighthouse at dusk and save it to my Ideas folder."
  • "Find my Costco receipts from January."
  • "Show me all my warranty documents."
  • "Move the washing machine manual to the Appliances folder."
  • "What receipts do I have for tax deductions?"
  • "Save these photos into my Medical folder." (uploads the attached images for processing)
  • "That PDF is the manual for my cordless drill — file it against that item."
  • "Add these three photos as pages of the drill manual."
  • "Read page 11 of the manual and tell me the torque setting."
  • "Those three photos are all the same manual — combine them into one document."
  • "Do I have the warranty for the dishwasher?"
  • "Detach that receipt from the lawnmower, it belongs to the hedge trimmer."
  • "Did that upload finish?" (checks the upload job's status)

#Views (Folders)

Organize your items and documents into folders and views.

Folder discovery and folder-create/move results include full paths from Home, so two folders called Marketing can be distinguished. Use the folder's real ID and full ancestry when choosing a destination. “Home › Marketing” means the top-level Marketing folder. If a name is ambiguous, show the candidate paths before writing; confirm the result using its returned folder_path_label.

Tool What It Does
list_views List folders at a given level of the hierarchy
get_view Get folder details and its first 20 documents
get_view_tree See the entire folder hierarchy at a glance, with item, document, and subfolder counts
get_view_members Get the union of items and documents in a folder
create_view Create a new folder. Idempotent — won't create duplicates
update_view Rename a folder, change its description, sort order, or thumbnail mode
reorder_views Batch update the sort order for multiple folders
generate_view_icon Generate an AI icon for a folder based on its name and contents. Costs credits
add_items_to_view Tag items with a folder — adds them without removing prior folder memberships
move_items_to_view Move items to a single destination folder
file_items File items into one or many folders in one call — by a filter (a category, everything in Home, everything in another folder, everything in any folder) or by explicit ids. "Put all the food in Home into the Food folder" is one group; "file everything in Home into the right folders" is one group per folder; "move everything out of every folder" is one group to Home with in_any_folder: true
remove_items_from_view Remove items from a folder
navigate_up Navigate to the parent folder

Example things you can say:

  • "Create a folder called Holiday Gifts."
  • "Move all the spices into the Spice Rack folder."
  • "How many items are in each folder?"
  • "Generate an icon for the Kitchen folder."

#Markdown Documents

Create, edit, and search your own written documents — notes, lists, reports, and more.

Tool What It Does
create_markdown_document Create a new markdown document with optional YAML frontmatter and folder placement
update_text_document Replace the contents of any TEXT document — note, .txt, .csv, .tsv. Send the complete new text, not a patch. Read it with get_document({include_source: true}) and edit source_text — not raw_text, which for a spreadsheet is a searchable summary rather than the file. Pass the source_hash you read as expected_source_hash — it is required, so a save someone made since you read it is never silently overwritten (SOURCE_HASH_REQUIRED without it, STALE_SOURCE if it changed: re-read and re-apply). A spreadsheet that doesn't parse is refused, not saved
undo_text_update Step a text document one state back. For a spreadsheet this restores the previous file, byte for byte
redo_text_update Step one state forward, undoing an undo
list_document_versions A note's saved versions, newest first — V number, when, and who saved it in the words to show (actor_label: "You", "You, via chat", "Grok, for you", "Shared with you: Kathy, via ChatGPT"). Every version from the last 90 days is kept, never fewer than the last 50 saves. "What did my Costco list say yesterday?" For a note shared with you, pass its owner_uid
get_document_version The full text of one saved version. On a note shared with you (owner_uid), lines someone else marked "Just me" are left out
save_document_version_as_copy Save a version as a new, editable document — "Costco Shopping List - V8" — in the same folder, leaving the original alone. For a note shared with you (owner_uid), the copy goes in your documents. "Get V8 of my packing list back as its own note"
get_markdown_document Retrieve a document's full content, parsed frontmatter, table of contents, and version info
list_markdown_documents List markdown documents, optionally filtered by folder, type, or tags
search_markdown_content Semantic + keyword search across markdown document content
read_document_rows Read a window of rows from a stored spreadsheet (.csv, .tsv, or Excel .xlsx) — pages through a large file the document text only partly shows. For a workbook it reads one sheet at a time, names the sheet it read, and lists the others
propose_import_mapping Work out what each column of a stored spreadsheet means before importing it, and recall the mapping you approved for that file layout last time
create_items_bulk Turn a spreadsheet — in your library, or attached in the current chat without saving it — into inventory items as ONE reversible operation — the server reads every row itself, so a file of hundreds of rows imports in one step; several columns (year, label, format…) can be kept together in notes; duplicates skipped, "undo that" removes the whole import
create_data_file Generate a file from content and save it to your library with a download link — Excel .xlsx, Word .docx, PowerPoint .pptx, or a plain .csv

Example things you can say:

  • "Create a packing list for my Hawaii trip."
  • "Make a spreadsheet of everything expiring this month." → create_data_file
  • "Give me a CSV of my inventory I can import elsewhere." → create_data_file
  • "How many rows are in that spreadsheet, and what's in row 500?" → read_document_rows
  • "Add everything in that spreadsheet to my inventory." → propose_import_mapping then create_items_bulk
  • "What sheets are in that workbook? Read the Q2 one." → read_document_rows (it reports every sheet, and takes a sheet by name or number)
  • "Write a summary of what I bought this month."
  • "Update the recipes section of my meal plan."
  • "Find my notes about home insurance."

#Skills

Skills are reusable SKILL.md playbooks that orchestrate your inventory, documents, tags, location, and beacons. The first-party set — curated by ambientChat — is available to everyone (read-only), and you can author your own on any Standard-or-higher plan. The same skills work in Claude and ChatGPT when you connect ambientChat there.

Tool What It Does
list_skills List the skills available to you — first-party plus your own — with name, slug, and description
get_skill Load a skill's full instructions (SKILL.md) by slug, ready to follow
create_skill Save a new skill of your own — name, slug, when-to-use description, and markdown instructions (Standard+)
update_skill Edit one of your own skills — its instructions, name, description, or tags. First-party skills stay read-only; creating your own with the same slug overrides them for you

Example things you can say:

  • "What skills can you run?"
  • "Use the Auto-Filer skill on this document."
  • "Run Renewals & Warranty Watch."
  • "Save what we just did as a skill called Weekly Fridge Audit."
  • "Update my trip-packer skill to always include chargers."

#Conversations & History

Your Q&A history — every question you've asked and answer you've received.

Tool What It Does
journal_conversation Save a Q&A exchange to your history for searchable recall
recall_history Search your prior conversations — past Q&A, advice given, topics discussed. Filters by channel (voice/text/mcp), platform, session, or date range
list_conversations Browse your full Q&A history in date order, with paging. Use for "show my recent conversations"; use recall_history to find a specific topic
trash_conversation Move a conversation to trash (soft delete)
trash_conversations Bulk move multiple conversations to trash
restore_conversation Restore a conversation from trash

Example things you can say:

  • "What did we talk about last week?"
  • "Recall the recipe advice you gave me yesterday."
  • "What questions have I asked about insurance?"
  • "Delete all conversations from March."

#Change History & Undo

Review and revert recent changes to your inventory. Every edit is tracked and reversible. Moves between folders, new items, and items sent to or restored from Trash are tracked too, and read in plain words ("Moved from Hall to Loft"). An edit or a move can be undone, as long as nothing has changed that field since. If something has, the undo says so instead of overwriting the later change. To bring back a trashed item, use restore_item.

Tool What It Does
recent_item_changes See recent changes to inventory items — old and new values for each edit, plus moves, creates, trashes and restores, each with a readable summary
list_item_location_history Where an item has BEEN — its location trail, newest first: manual "Last seen" sets, the stamp made when its tag was linked, and live tag sightings. Try: “where has my tape measure been?”
list_item_temperature_history A tracked item's temperature trail from its tag's telemetry — minute points, newest first. Try: “what temperature is my tape measure at?” or “did the cool-box get warm overnight?”
get_asset_journey The chain of custody for one or more tracked items over a date range: where they went, grouped by venue and zone, how long they stayed, the temperature along the way, and the events it was built from — each stop says whether it was observed (tag heard, scan) or entered by a person. Try: “where has this crate been this week, and did it stay cold?”
create_journey_report Publishes a tracked item's journey as a live page anyone with the link can open — venues and zones with time spent, temperature, a map, and how each stop is known — saved as a document and shared through the normal public link. Positions are rounded to about 10 km unless you ask for exact ones. Up to 25 items, named or as a folder the page follows live — several items show as one row each on a shared clock. Try: “make a journey report for the crates in my Demo Kit folder.”
undo_item_change Undo one edit or move of an item, putting back what it replaced. It refuses if that field has changed since. recent_item_changes says which changes it would accept (can_undo)
get_request_receipt Whether an item or document edit the ambientChat app sent with an action key (act-<uuid>) ran: landed, not_applied (it ran but the record was gone, so nothing changed), running, or unknown when no receipt is kept for it. Only landed means it took effect. Receipts are kept at least kept_hours (24). Only the app's own edits carry such a key, so one an AI app made reads unknown; your own account only
recent_bulk_changes List recent bulk inventory operations — useful after a mass update
undo_bulk_change Revert an entire bulk update, or a filing by file_items / move_items_to_view, in one call — every item the operation touched goes back (a filing returns each item to the folders it was in). Anything it left alone is in skipped, each with a code saying why

Example things you can say:

  • "What did you just change?"
  • "Where has my tape measure been?"
  • "Where was the drill before the garage?"
  • "Undo that last edit."
  • "Show me the recent bulk changes."
  • "Revert the bulk update you made to my kitchen items."

How undo works: Every change is logged with before/after values. You can undo individual edits or entire bulk operations. Undo is available from any surface — in-app chat, voice, Claude Desktop, or the web interface.


#Sharing

Share your items, documents, conversations, and folders — with a person by email or with a whole group — and manage what others have shared with you. See the Sharing guide for the full walkthrough.

Share something out:

Tool What It Does
share_with_email Grant a person live access to an item, document, folder, conversation, or your own Desktop AI Mac (home_node) by email — read-only or read-write. Sends an invitation if they have no account yet. A Mac takes one of its 5 seats (refused with home_node_seats_full when they're taken) and never expires
share_taste Share a slice of your taste — only the kinds you pick (kinds, e.g. ["clothing"]) — with one person (recipient_email) or any of your connected AI apps (app_keys, e.g. ["claude", "grok"] — one share each, all made together or none; app_key still works for one). The reply lists each share in grants. They read a short summary written from that slice alone: nothing about any other kind, and never your interests. include adds dislikes or finished likes (default: likes only); level: "signals" adds the liked things by name. Call with preview: true first: it returns the exact text they will read and shares nothing. Every taste share ends (expires_in_days 30, 60 or 90; default 90); stop it any time with revoke_share_grant. In ambientChat's own chat it always asks first, showing that text. "Share my clothing taste with Mathieu"
get_shared_taste Read taste someone shared with you — or, for a connected AI app, the slice you shared with it: each share's kinds and its summary. Refused once the share is revoked or has ended, and for an app whose access you paused. Each read shows in the owner's sharing list ("Mathieu read your clothing taste"). "What's Mathieu's clothing taste?"
share_with_group Share something with an entire group at once. Choose how long it lasts (expires_in_days: 30, 60, 90 or never; 30 by default), as when sharing by email. Try: “share the camping folder with Family for 60 days”
share_item Share an inventory item
share_document Share a document — includes title, type, and summary
share_conversation Share a Q&A answer
share_view Share a folder — cascades to its sub-folders and contents
list_my_shares See everything you've shared and with whom
renew_share_grant Give an in-app share more time — "renew Ann's access". Never shortens: a share that already has longer is left as it is
renew_share The same, for a share you sent by email or as a link — so the link already in someone's inbox keeps working. Also only ever extends, and a public link has no expiry to renew
update_share_grant Change a share's access (view ↔ edit) or set an exact length, which can be shorter. Returns the previous values so the change can be undone
revoke_share_grant Remove a share — access ends immediately
revoke_share Remove an emailed or link share — the counterpart of revoke_share_grant for links
share_list_link Make (or get back) the public link to a list, and choose whether whoever holds it can tick lines off — with no account, and nothing else. can_tick: false makes the same link look-only. Lists only. Try: “text Kathy the Costco list so she can tick it”
resend_invitation Re-send a pending invitation to someone who hasn't signed up yet

Groups:

Tool What It Does
create_group Create a group to share with several people at once
invite_to_group Invite a person to a group by email
accept_group_invitation Accept an invitation to join a group
list_my_groups List the groups you belong to
get_group_members See who's in a group
remove_group_member Remove a member — ends their access to the group's shares
set_group_member_role Make a member an admin of the group, or an admin a plain member. Admins can invite, remove and promote; a group always keeps at least one admin, so the last one cannot be demoted or removed until someone else is made admin
delete_group Delete a group

Shared with you (and mounting):

Tool What It Does
list_shared_with_me See what others have shared with you
get_shared_resource Open a specific resource someone shared with you
list_shared_folder_contents List the contents of a shared folder. With kind (item, document or folder) it pages one kind at a time like your own folder, most recently changed first unless sort_by says otherwise — pass next_cursor back as cursor until it is null; total is null in this mode. Without kind, the whole folder up to 2,000 documents and 2,000 items; total is its real size and truncated says when the list stopped short
hide_shared_with_me Remove a share from your list (hides it for you only; owner's copy is untouched)
mount_shared_resource Add something shared with you (or with Everyone) into your own library, so it sits inline alongside your own content — like "Add to My Drive". The master stays with its owner, so you always see their latest version
unmount_shared_resource Remove a mounted shared resource from your library (reversible; doesn't delete the owner's content or revoke your access)
list_mounts List the shared resources you've added to your library
list_mounted_contents Resolve the shared resources you've mounted into a folder into their live data + provenance, so they show inline alongside your own content

Example things you can say:

  • "Share the olive oil with my wife, read-only."
  • "Share my Recipes folder with the family group."
  • "Who have I shared the Warranties folder with?"
  • "Stop sharing that receipt with the repair shop."
  • "Add the shared Skills to my library."

#Connected Services

Connect outside tools (GitHub, Gmail, Calendar, Drive, Slack, Linear, Mercury, or any MCP server by URL) so the assistant can use them in chat alongside your own context. See the Connected Services guide.

Tool What It Does
list_mcp_servers List the services you've connected, with status and tool counts
add_mcp_server Connect a new MCP server by URL — OAuth handled where the provider supports it
update_mcp_server Edit a connection's settings or stored token
remove_mcp_server Disconnect a service — its tools leave chat and its access is revoked

Example things you can say:

  • "What services do I have connected?"
  • "Connect my GitHub."
  • "Match my Mercury transactions to the receipts in my library."
  • "Disconnect Slack."

(Connecting outside tools into ambientChat is the opposite of Context Sharing — the get_context_sharing tool and the context_sharing setting below, which control what ambientChat shares out to Claude, brands, and partners.)


#Tool Permissions

Control what the assistant can do without asking — Anthropic-style per-tool cards. Every tool is Auto (runs automatically), Ask (approval prompt first), or Blocked (never runs). Writes to connected services ask by default; when a prompt appears you can tap Always Allow to stop being asked for that tool. Manage everything under Control → MCP Servers → Tool Permissions (iOS and web). Safety tools like emptying trash or deleting your account always ask and can't be changed.

Every tool has two settings: ambientChat chat (Auto, Ask or Blocked) and Outside AIs — ChatGPT, Claude and other AI apps you connect (Auto or Blocked). Outside AI apps can't ask you first, so for them a tool is either allowed or blocked. A tool you had blocked before the two settings existed starts blocked in both. On the web, Tool permissions is one click from Control, on the MCP Servers card.

You don't have to set tools one by one: ask ambientChat's chat, e.g. "block anything that deletes for outside AIs" or "stop ChatGPT changing my inventory". It shows the changes as a card under its answer, and nothing changes until you tap Apply. Outside AI apps can never see or change your tool permissions: only you can, in the app or by asking ambientChat's own chat.

Tool What It Does
list_tool_permissions List every tool — built-in and from connected services — with its approval status for ambientChat's chat and for outside AI apps. ambientChat's own app only
update_tool_permission Change a tool's policy to approved, ask, or blocked, for ambientChat's chat or for outside AI apps — or several at once with changes, all or none (asks you to confirm the change). ambientChat's own app only
propose_tool_permissions Suggest a set of permission changes — e.g. "block anything that deletes for outside AIs" — as a card you apply with one tap; nothing changes until you tap Apply. ambientChat's own app only

Example things you can say:

  • "Which tools need my approval?"
  • "Stop asking me about Linear writes."
  • "Block web search."
  • "Make save_issue ask again."

#Trash & Restore

Manage your trash across all content types.

Tool What It Does
list_trash See everything currently in the trash, ordered by most recently trashed
restore_trash Generic restore — works for items, documents, and conversations without needing to know the type. Takes one id, or several at once for "restore everything"
empty_trash Permanently delete ALL trash entries. Optional scope: "item", "document", or "all"
purge_trash Permanently delete the entries you name, leaving the rest of the trash alone. Purging a trashed folder takes its contents with it

Example things you can say:

  • "What's in my trash?"
  • "Restore the last thing I deleted."
  • "Empty the trash."

#Search & Discovery

Cross-cutting tools for finding information and getting answers.

Tool What It Does
web_search Search the web for current information — used when you need real-time data, recent events, or fact verification
query_expert Ask another AI model for a second opinion mid-conversation. The answer is credited to the model that gave it — "Gemini thinks…" — never passed off as your agent's own. To review a document, pass its id in include_documents: any of your documents goes to the expert in full, whether or not it matches the question (so does one someone shared with you, credited to its owner), and naming documents makes it a review of those documents alone — no personal context and no tools (context: "none"; pass context: "auto" to include your context too). Without named documents the expert sees your ambientChat context, and the result's documents list says which large documents the answer used and which were held back, with the token cost; exclude_documents leaves one out. Long answers never get lost: if the expert needs more than about 40 seconds — a reasoning model reading a long document can take minutes — the call returns status: "running" and a job_id instead, and the answer is saved either way. Grok, which never collects a job_id and gives up on a call after a minute, is held for the answer itself for up to 50 seconds; asking the same question again within 10 to 20 minutes is answered from the first run, not a new paid one. A retry is never charged twice: sent again with the same _meta.idempotency_key, it is answered from the first answer (replayed: true, credits_used: 0), or with the first call's job_id if that is still running; the same key with a different question is refused (REQUEST_ID_REUSED)
get_expert_answer Collect the answer to a query_expert call that returned status: "running". Your agent calls it with the job_id every poll_after_s seconds (10 s at first, up to 30 s on a long run) until the answer is completed (same fields as query_expert) or failed. Nothing is lost by the wait: the answer is also saved to your conversation history. A failed answer says why (failure_reason: refusal, max_tokens, timeout, rate_limited, credits, too_large, provider_error, internal, stopped, save_failed, no_text). It also returns any words written (partial_answer, incomplete) and names suggested_fallback_model, a model of the same lab to offer the user, never called on its own
get_user_context Assembles relevant context for your query — bio, nearby beacons, focus item details
present_records Shows your items, documents, folders or saved answers as cards — each with its real name, picture and a link that opens it in ambientChat. Items and documents also say where they are filed ("in Garage", "at Home"), exactly as the app's own cards do. Try: “show me my power tools”
resolve_scanned_code Says what a scanned code is — the same answer the phone's scan card gets. A barcode or ISBN becomes the product or book; a web address becomes the page (with your own title and picture if you kept it). A Bluetooth tag's printed label — its address, like C3:00:00:51:CC:E7 or c3-00-00-51-cc-e7 — comes back as a Tag with its address in the one standard spelling, ready to look up with list_asset_tags. Try: “what is this code: 036000291452?” or “what is c3-00-00-51-cc-e7?”
update_beacon Renames a registry beacon or tag, or changes how it BEHAVES, on the central record every app reads: its name (a renamed tag's name also updates on the item it tracks), alert mode (silent01, popup01, ack01), alert text, context string (what the AI is told whenever it is detected), context prominence (1–3) and live/paused. Only the fields you pass change; the previous state is kept in the record's history. The beacon's owner, a brand admin for its brand, or staff may change it. Try: “pause the lobby beacon”, “make the Study beacon high prominence” or “rename my Keys tag to Front door keys”.
delete_beacon Deletes a registry beacon or tag: it leaves the registry, so it stops appearing in the owner's list, on every device and in their AI context; its last state is kept in the record's history. An item a deleted tag tracked stops being tracked, and the tag can be set up again by scanning its label. The owner, a brand admin for its brand, or staff may delete it, and you are asked to confirm first. Try: “delete my old wallet tag”.
release_tag Releases a tag you own, so another account can set it up as theirs by scanning its label: it stops being yours, its links and the tracking on its items are removed, and if it is in a Zone it leaves the Zone too. Its name, notes, alert text and group are cleared, so the next owner sees none of them. It stays in the registry. Only its owner may release it, and you are asked to confirm first, with the Zone it will leave named. Try: “release my old bike tag”.
link_beacon Links a registry beacon/tag to your inventory items or documents. On a FIXED beacon (yours, or any for staff) it ADDS an exhibit (up to 8 per beacon); unlink: true with a kind and id removes just that one, and kind none clears them all. On a Bluetooth TAG (yours, or any for staff) it starts asset tracking — the tag follows one item, replacing its previous one, and keeps any documents on it. Try: “link the lobby beacon to the Rodin sculpture” or “track my drill with the Minew tag”. A link to an item or document that does not exist is refused, and a write drops any of the beacon's links whose targets are gone.
list_asset_tags Lists the Bluetooth asset tags you own (tags only — never fixed Zone beacons), each with the item it currently tracks or free, its battery state, and — when the battery reads unknown — why. A tag the app set up over Bluetooth also says so (provisioning_state: active) with the firmware, model and profile it was given; a tag you already had stays listed through a re-flash, with the state saying how that went; a tag you have never had is not listed while it is being set up or after a failed first attempt. Try: “which of my tags are free?”, “which of my tags did the app set up?” or “why is my tape measure tag’s battery unknown?”
list_beacons Lists the FIXED beacons you own (Zone beacons and exhibits — never mobile asset tags), each with its battery state, when that reading was taken, when it was last heard, what it is linked to (linked_refs), and — when the battery reads unknown — why. Try: “what's linked to my Pantry beacon?”, “which of my beacons need batteries?” or “is the garage beacon dead?” — a flat beacon battery looks exactly like bad signal or a detection bug, so it is worth ruling out first
reserve_device_identity Claim and HOLD an iBeacon identity (UUID + Major + Minor) for a Bluetooth device you are setting up, so nobody else can be given the same number while you work — and get back the write plan: which slots the phone will program at what interval and power, from the standard profile for a tag or a beacon. Leave the network out and the device is placed on the fleet's own. Ask again with the same MAC address and you get the same identity and the same plan back rather than burning a second one, so a retry after a failed write is safe; prior says what the row was before the call (new, reserved, failed or active), which is how the phone knows to read a device's current values before overwriting one that was set up before. The hold lapses after 24 hours if the setup is abandoned; a device that is live keeps its identity for good. The plan's password never travels to an agent — only the phone's own channel carries it. Try: “reserve an identity for the tag with MAC AA:BB:CC:DD:EE:FF”
create_venue Create a place — your home, your office, a store. It's personal by default: only you can see it, and it never collides with anyone else's place of the same name. Pass visibility: "group" to share it with a group you belong to, or "public" to list it for everyone (that one needs a brand-admin or superuser role). Try: "add my house as a place"
update_venue Rename a place or move its coordinates. You can edit any place you created; editing never changes who can see it
create_zone Add a room or area inside one of your places — "Kitchen", "Meeting Room 2". Zone names must be unique within their venue; a duplicate is refused and names the existing one so you can use that instead. A zone is visible to whoever can see its venue
assign_beacons_to_zone Point one or more registered beacons at a zone — this is what puts a beacon on the map. Needs a brand-admin or superuser role, because a beacon is shared hardware that may already belong to someone else's place

Example things you can say:

  • "What's the current price of a KitchenAid mixer?" (triggers web search)
  • "Ask Gemini what it thinks about this recipe." (query expert)
  • "What do you know about me?" (user context)
  • "Add my house as a place, and put a Kitchen zone in it." (creates a personal venue + zone)

#Settings & Profile

Read and update your profile and app settings — all manageable via voice or chat.

#Profile

Tool What It Does
get_profile Get your profile — name, nickname, email, photo, timezone, home/work locations, and bio
update_user_setting (key profile) Update profile fields — first name, last name, nickname, hail name, home/work location, bio, and taste_summary (your own wording of your taste summary; null makes a fresh one from your 👍/👎 and interests, replacing your wording — agents ask first)
upload_profile_image Upload a new profile photo from image data
reimagine_profile_image Re-render your profile photo as a claymation figurine portrait

#Identity & Roles

Tool What It Does
list_users_directory Read-only staff directory with role-scoped accounts/publication contacts, search, state filters, complete matching counts and paging. Website Administrators see ambientChat newsletter contacts; publishers see Mr. Beacon; superusers see the unified directory. Does not import, enroll, export or send. Legacy Mailchimp history remains unreconciled.
get_identity Who you are and what you can do — your uid, email, role (superuser / brand-admin / brand-user / enduser), whether you're a superuser, and additive website roles/capabilities
set_user_role (Owner only for website grants) Change a base role or additive website_roles by uid/email. Omit role to preserve the base role; an empty website_roles list revokes delegated website access. Audited; unrelated Auth claims are preserved.
website_content Read, save, publish or roll back website content. Website Administrator manages main; Mr. Beacon Publisher manages existing episode:<id> entries. Use the returned revision for writes and historyCursor for earlier publications. Private drafts, explicit publication, scoped permissions and experiment content locks apply equally to HTTP and MCP.

#Interests ("How I Roll")

Tool What It Does
list_interests List your interests / "how I roll" tags — free-text preferences that personalize AI responses
add_interest Add an interest tag (e.g. "vegetarian", "loves hiking")
remove_interest Remove an interest by id

#AI & Scanning

Tool What It Does
get_ai_model_pref Get your default AI model preference
update_user_setting (key ai_model) Set your default AI model — validated against the model registry
list_available_models The models you can pick right now. If you have Desktop AI — a model running on your own Mac — it is listed first by your Mac's name and the model it runs (“Steve's Mac – Qwen 3.6 35B”), with whether it's online and, if not, why (e.g. “Offline since 2:02 PM”). It costs nothing: questions your Mac answers use no credits. An offline Mac is never swapped for another model. Try: “is my Mac online?”, “which models can I use?” or “which model is cheapest?”
register_home_node Superuser only (Desktop AI early access). Register a Mac for an account so its Desktop AI appears in that account's model list. Returns the Mac's connection key once
start_home_node_setup Desktop AI. Set up Desktop AI on your own Mac: names the Mac and gives you one command to paste into Terminal on it. The command works once and expires in 30 minutes; ask again for a fresh one. Pass the Mac's node_id to set it up again — that keeps the model it runs, while it still fits the Mac's memory. Otherwise ambientChat picks the model for the Mac's memory: 48 GB or more gets Qwen 3.6 35B (64k) (home-qwen-3.6-35b-a3b-q3-64k, conversations twice as long), a smaller Mac Qwen 3.6 35B. A new Mac, or one too small for the model chosen, is picked for the memory it measures when it connects. ram_gb is the memory the Desktop AI app measured; it counts only until the Mac reports its own. list_available_models says what became of a Mac's latest setup code (setup_state: pending, claimed or expired). The answer says which model and why (model_build_reason: kept, picked_for_ram, default or requested; model_build_note in words). Superusers can pass model_build_id to choose another vetted model (e.g. home-gemma-4-26b-a4b-q4, Gemma 4 26B), or a build being tested on a 64 GB Mac and not yet scored: home-qwen-3.8-27b-q6-64k (Qwen 3.8 27B). Try: “set up Desktop AI on my Mac mini”
remove_home_node Desktop AI. Remove one of your Macs for good — an old Mac, or the duplicate a reinstall left behind. It leaves your model list on every device, can no longer connect as your Desktop AI, and a default that named it is cleared. Asks you to confirm in chat; can't be undone (set the Mac up again to use it). Try: “remove my old Mac mini from Desktop AI”
set_private_ai_access Desktop AI (alpha). Switch Desktop AI on or off for your own account — “Try Desktop AI”, open to every account with no invitation (list_available_models says private_ai_access available while it is off). Switching it off also clears your default model if it was one of your Macs (default_cleared), so chat moves to a cloud model. Superusers can also give another account Desktop AI, or take it away, by email. Try: “turn on Desktop AI”
email_desktop_ai_setup_link Desktop AI (alpha). Email yourself the link to set up Desktop AI on your Mac (Control → Desktop AI on the web, where the Mac downloads the app) — for when you're on a phone or a PC. It only ever goes to your own account email. The link is always returned too: from the phone or web button it's emailed, and when an AI assistant can't send mail it gives you the link directly. Try: “send me the Desktop AI setup link”
share_with_group (a Mac) Desktop AI (alpha). Share one of your Macs with a group (resource_type home_node, resource_id the Mac's node_id): its members, up to 5 people, can use it as a model. A bigger group is refused (home_node_seats_full), and anyone who joins later waits for your OK. revoke_share_grant stops a group using it. Try: “share my Mac with Family”
list_home_node_people Desktop AI (alpha). Who can use one of your Macs: each person seated or waiting for your OK, the groups it's shared with, and how many of its 5 seats are taken. Never what they asked. Try: “who uses my Mac?”
approve_home_node_seat Desktop AI (alpha). Let one person in a group you shared your Mac with use it. Only on your say-so; asks you to confirm. Try: “let Alex use my Mac”
remove_home_node_seat Desktop AI (alpha). Stop one person using your Mac without removing them from the group; their seat frees. Asks you to confirm. Try: “stop Alex using my Mac”
get_desktop_ai_guest_usage Desktop AI (alpha). How much each person used your Mac, by month and by kind (questions, web checks, images, expert opinions). Never what they asked: no content is ever recorded. Try: “how much does Sam use my Mac?”
answer_home_node_default_offer Desktop AI (alpha). Answer the one-time “Use Ana's Mac as your default?” for a Mac someone shared with you. Either answer is final; asks you to confirm. Try: “yes, use Ana's Mac as my default”
acknowledge_home_node_disclosure Desktop AI (alpha). Record that you've read a shared Mac's first-use notice (“Your questions are answered on Ana's Mac. We never store them there, but only use a Mac whose owner you trust.”), so no device shows it again. Only after you've read it; asks you to confirm
get_home_fleet_versions Superuser only. Which Desktop AI version each Mac runs and where it was built (an official build, or a local build or experiment), whether it follows main (dev) or tagged releases, who is behind, and which Macs went back after a failed update — and why. Try: “which Macs are behind?”
get_home_build_report Superuser only. How each Desktop AI build performed over the last days: calls, failure rate, how fast the first word arrives, cold starts, reading and writing speed, and how much of each question the Mac could reuse from its memory of the last one instead of reading it again. Local experiments are listed apart and never counted toward a release. Also on the web admin: Diagnostics → Desktop AI. Try: “did that change make the Mac faster?”
set_home_agent_rollout Superuser only. Move the Desktop AI release rollout: founder Macs first, then a canary share, then everyone — or halt it. Widening is refused if the new version fails more often or answers more slowly than the one it replaces
get_tag_location_rollout Superuser only. Who has server-side tag location: which accounts are in shadow (the server works out where each tag is and compares it with the zone the item says, but never moves an item) or on, which matcher runs, and who last changed it and when. Try: “who is in tag location shadow?”
set_tag_location_rollout Superuser only. Move accounts in or out of server-side tag location, by uid or email: off, shadow or on. Putting anyone on needs confirm: true; until the server is allowed to move items, on behaves exactly like shadow. Roll back with mode off. Other servers pick up a change within a minute. Try: “put my account in tag location shadow”

#App Behavior

Tool What It Does
get_app_behavior_prefs Get app-behavior preferences — haptic feedback, sounds, keep source photos
update_user_setting (key app_behavior) Update haptic feedback, feedback sounds, or the "keep source photos" toggle
get_background_trails_pref Get your background location trails master switch state
update_user_setting (key background_trails) Toggle background location trails on or off

#Save Destinations

Tool What It Does
get_save_preferences Get your sticky save-destination — the folder new scans/items default into
update_user_setting (key save_destinations) Set or clear your default save folder

#Notifications & Context Sharing

Tool What It Does
get_notification_preferences Get beacon-category notification preferences — which categories are disabled
update_user_setting (key notifications) Toggle notifications for a single beacon category, or bulk enable/disable all
send_test_access_alert Send a test of the alert you get when an AI app reads or changes your documents and context, to every device registered on your account, and report device by device whether Apple (or Google) accepted it. Try: “send me a test alert” or “am I getting access alerts?”
get_beacon_snoozes Read the beacons whose alerts you've snoozed, and until when. A snooze is yours alone: it follows you to every device and never silences the beacon for anyone else. Snooze or un-snooze with update_user_setting, key beacon_snoozes. Try: “snooze the lobby beacon for an hour” or “which beacons have I snoozed?”
get_venue_invitation_settings Read the venue-invitation settings: the master switch plus the venues you've told "don't show this invitation again" — synced to your account
update_user_setting (key venue_invitations) Turn all venue invitations on/off, or opt out of (and back into) one venue's tour invitation — every change reaches every device
get_context_sharing Everything the Context Sharing screen shows — the master switch, every connected AI app with its own switch, and the access-alert switches
update_user_setting (key context_sharing) Turn the master context-sharing switch and the access alerts on or off (pause one AI app with set_connected_app_sharing)

#Account Management

Tool What It Does
get_account_deletion_precheck Pre-flight check — shows data counts, subscription status, warnings, and your latest export (latestExport) or whether one is still being made (exportInProgress), so the AI can offer — never require — saving a copy first
request_account_deletion Request account deletion. Creates a confirmation with 5-minute expiry

Example things you can say:

  • "Switch my default AI to Claude."
  • "Turn off haptic feedback."
  • "Turn off location trails."
  • "Stop sharing context with Amazon."
  • "Give me a claymation profile picture."

#Privacy

Control how your ambient data (location, beacons) is shared with AI calls.

Tool What It Does
get_privacy_preferences Get your ambient-data privacy preferences — location precision per AI call type
update_user_setting (key privacy) Update location precision settings for different AI call types
list_connected_apps Every AI app connected to your account — one row per app, with whether it may read your context, when it last did, and what it read recently
set_connected_app_sharing Pause or resume one app's access without touching the others. Asks you first; an outside AI app can pause but never switch an app back on
disconnect_connected_app Sign one app out on every device — it can't read anything until you connect it again from that app. Asks you first

Example things you can say:

  • "What are my privacy settings?"
  • "Don't share my exact location with AI — use city-level only."
  • "Which AI apps can see my data?"
  • "Pause ChatGPT." · "Disconnect Muse."

#Voice Preferences

Control how the assistant speaks to you.

Tool What It Does
get_voice_preferences Get your voice/TTS settings — voice name, style, auto-speak, playback speed, and whether the Radar's Zone guide is on (tour_listen)
update_user_setting (key voice) Update voice name, style, auto-speak toggle, playback speed, or the Zone guide (tour_listen) — "stop the tour guide"

Example things you can say:

  • "Switch to the Shimmer voice."
  • "Speed up your speech."
  • "Stop auto-speaking responses."
  • "What voice are you using?"

#Settings Catalog

Browse all available settings and their current values.

Tool What It Does
list_user_settings Read every settings domain you have in a single structured payload
list_setting_dimensions Read the catalog of settings the AI can manage on your behalf — each names the tool that changes it (usually update_user_setting with its setting_key)
update_user_setting Change ANY setting: pass the domain's setting_key plus the fields to change — one write tool for every settings domain

Example things you can say:

  • "Show me all my settings."
  • "What settings can you change for me?"
  • "Turn haptics off." (one update_user_setting call — same for any setting)

#Subscription & Usage

Check your plan, credit balance, and usage history.

Tool What It Does
get_subscription Get your subscription summary — tier, period dates, credits used/remaining, exemption status
get_usage_state Get current usage and billing summary — tier, credits, exemption status
get_billing_events Get itemized billing history — individual AI requests with credits consumed, model used, and token counts
grant_credits Superuser only. Add complimentary credits to a user's current period (plan and reset date unchanged; expires at reset). The note is shown to the user

Example things you can say:

  • "How many credits do I have left?"
  • "What plan am I on?"
  • "Show me my usage this month."
  • "What's my most expensive AI request?"

#Export & Import

Move your data in and out of ambientChat.

Tool What It Does
print_resource Print an item, document, or saved answer — renders the branded PDF and (in the ambientChat app) opens the print dialog automatically
export_user_data Start an export of all your data as a ZIP archive — inventory, documents, conversations, profile, and your settings and AI rules in one readable rules file. A full export becomes your backup and replaces the previous one; a CSV export doesn't
get_export_status Check if your export is ready and get the download URL
list_export_jobs Your recent exports and your one current backup — with a fresh download link and its size
get_import_upload_url Get a secure upload URL for importing data
get_import_status Check how an import is going, and what it restored when it finishes
list_import_jobs Your recent imports — one still running, what the last one restored, or why it failed
import_user_data Import data from an uploaded archive — preserves original IDs and reconciles links

Example things you can say:

  • "Print this document."
  • "Print my coffee maker."
  • "Export all my data as a ZIP."
  • "Is my export ready?"
  • "Import data from the file I uploaded."

#Onboarding & Help

Get started and find help.

Tool What It Does
get_ambientchat_welcome Get your personalized welcome — tiles, greeting, suggested prompts, and next best action
get_help Look up how ambientChat works: returns the section of the Don't Panic guides that answers your question (a few hundred words, with the guide, the section and a link to it), plus pointers to nearby sections
show_screen Point you to the app screen where a how-to answer's steps happen, named from the app's list of screens (for example bluetooth-tags, desktop-ai, documents, model-picker). Inside ambientChat it shows a Take me there button under the answer (on the web, a screen that only the phone has gets a line saying it's done on your phone instead). An outside AI such as Claude or ChatGPT gets a url and a sentence to pass on. The link is an ambientChat web address (…/screen/bluetooth-tags, or …/coach/<flow> from coach_steps): on a phone with the ambientChat app it opens the app at that screen; anywhere else it opens the web page, which starts the walkthrough for a /coach/ link (signing you in first if needed), or says the screen is on the phone. The sentence reads "Open Documents in ambientChat: …", or, for a screen that only the phone has, "This is done in the ambientChat app on your phone, under Bluetooth tags: …" with the link Asked for Control in general with a reason that names one of its sections — "choose which AI apps can access your context in Context Sharing" — it opens that section (here context-sharing, Control scrolled to Context Sharing) rather than the top of Control, and returns that id as destination. Instead of a screen it can name one of your own things — an item, a document or a folder — when you ask how to get to it ("how do I get to my Music folder?"): the button is titled with its name and opens it (in an ambientChat app older than this one, the answer links it instead). An outside AI sends it to you as a card, as it does a screen ("Claude wants to show you Music"), and gets a link to pass on (…/thing/view/<id>): on a phone with the ambientChat app it opens the app on that thing; anywhere else it opens its page in the web app. If your phone has an ambientChat app older than this one, it gets no card for a thing, and the AI passes on the thing's web link instead, which opens in your browser. A folder of documents opens on Documents, by the card or by the link. Asking where something is ("where's my Music folder?") is a question about your stuff, answered with links and cards, and gets no button. It can also name one control on that screen in the words you see there — "Notify when accessed" — and then the button opens the screen with that control circled; if the control sits inside something that has to be opened first (a beacon's row, a closed section), it circles that first, unless it is already open. On a screen that lists your own things — the AI apps in Context Sharing, your models, your beacons, your muted venues — it can name one by its name: "help me turn off sharing with Grok" circles Grok's switch, and if no row by that name is there, it circles the list. A control it can't find on any other screen isn't refused: the screen still opens, with nothing circled. When it does find the control, an outside AI's sentence names it ("…Then tap Notify when accessed."); the circle itself appears in the ambientChat app, including when an outside AI sends the screen to your phone or browser as a card: Open takes you there with the control circled (an app older than the control opens the screen without it). It opens a screen without a tap only when you asked to be taken there, and a screen name it doesn't know is refused with the list of ones it does
coach_steps Walk you through a how-to on the real screen. It names the screen and up to six steps, each one a control from the app's own list (for example the Share button, the Recipient box, They can), with a short hint written for what you are doing — "Type Mathieu's email here". Inside ambientChat it shows the same Take me there button; tapping it opens the screen and shows each hint on its control, one at a time, while you do the task, and changing that field moves on to the next. × stops it. A control or screen it doesn't know is refused, and a step for a control that only one surface has is refused on the other (a phone-only control on the web, a web-only one on the phone). An outside AI such as Claude or ChatGPT gets a url and a sentence to pass on, as with show_screen
get_coach_status Follow up on a screen an outside AI sent to you. When an ambientChat app of yours can show the request as a card, show_screen (for a screen, or one of your own things) and coach_steps hand it there and return sent_to_phone, sent_to (phone or browser) and a request_id; nothing opens until you tap Open. It goes to whichever is in front most recently: the ambientChat app on your phone, or ambientChat in a web browser you are looking at (a browser counts only while it is in front, and only for a screen — and steps — the web has, or one of your things; a phone-only screen always goes to the phone). The card reads "Claude wants to show you Bluetooth tags" — Nothing changes until you choose. — with Open and Not now (an app it can't name shows as An AI app). Open takes you there and, for coach_steps, walks you through it; for show_screen with a control, it circles that control, and from the link beside the card it circles it on whichever of your devices you open it on (an ambientChat app older than the control opens the screen without the circle, and the AI is told so). The AI passes that request_id here and learns how it is going: pending (the card is waiting), opened, step (which step of how many), completed, abandoned, declined (you tapped Not now) or expired (unanswered for 10 minutes). For one of your things it also returns which thing (thing). With nowhere to send it (an older phone app, no phone used in a day, and no browser in front), there is no request_id and the AI passes on the link instead
propose_field_values Fill in a form for you to check and save: an item's name and place, or a document's share recipient and access (view or edit). It writes nothing — inside ambientChat a Fill in button under the answer opens the form with the values filled in and ringed, and nothing is saved until you save. In other AI apps there is no form to fill, so it tells the AI to make the change with the regular tools instead (update_item and set_item_location, or share_with_email), which can be undone
get_onboarding_state Read your onboarding progress — tier, engagement metrics, dismissed tiles, activation status, and which once-only walkthroughs you have already seen (seen_walkthrough_ids, e.g. radar_intro for "Show me around Radar")
update_onboarding_state Update onboarding preferences — turn the welcome suggestions off (stop_showing) or back on (resume_showing), dismiss tiles, restore dismissed items, reset; mark a walkthrough seen (mark_walkthrough_seen) or bring it back (reset_walkthrough — "show me around Radar again"), or bring every one back at once (reset_walkthroughs — "show me all the tours again"; leaves dismissed tiles alone)
record_onboarding_event Record onboarding analytics events (used internally for A/B analysis)

Example things you can say:

  • "How do I scan a receipt?"
  • "Show me the getting started guide."
  • "How do I set up a Bluetooth tag?" (steps, then where to do it)
  • "Turn off the welcome suggestions."

#User Stats

Tool What It Does
get_user_stats Get counts of your content — active and trashed inventory items, documents, folders, and conversations

Example things you can say:

  • "How many items do I have?"
  • "Give me an overview of my account."

#Confirmation Tools

Safety gates for irreversible actions. Some operations create a pending confirmation that you must approve within a short time window.

Tool What It Does
list_pending_confirmations See actions awaiting your approval
confirm_action Approve a pending action
deny_action Reject a pending action
list_pending_chat_approvals See the Approve/Deny questions a chat thread is still waiting on — what the in-app prompt asked. Each expires about two minutes after it was asked; answer it in the app

Example things you can say:

  • "Is that chat still waiting for me to approve something?"
  • "What did that chat ask me to approve?"

list_pending_confirmations lists actions an outside app asked for, which confirm_action answers. list_pending_chat_approvals lists what a chat is waiting on, which is answered in the app.

Tools that require confirmation: request_account_deletion, purge_trash Tools that require failsafe confirmation: empty_trash


#Scan Regression Tools (Developer Only)

These tools are restricted to superusers for internal quality testing.

Tool What It Does
get_scan_regression_runs List recent scan-regression harness runs with summary metrics
get_scan_regression_run Fetch one scan-regression run with full per-fixture detail
run_scan_regression Trigger a scan-regression run against deployed code
get_scan_regression_request Check the status of a previously created scan-regression request
get_benchmark_runs List recent Desktop AI benchmark runs with pass rates, answer times and cost
get_benchmark_run Fetch one benchmark run with every question's answer and why it passed or failed, or one attempt's full transcript
run_benchmark Start a benchmark run for a model (superuser; takes 5-15 minutes)
get_scorecard_runs List stored Model scorecard runs (answer quality per production step) with pass rates and their 95% range (superuser; dev only)
get_scorecard_run Fetch one scorecard run with every attempt and the grader's reason, or one attempt's prompt and answer (superuser; dev only)
get_scorecard_comparison Compare 2-6 scorecard runs against the first, task by task, with the gap and whether it is inside the noise (superuser; dev only)
get_coach_scorecard Each model's coaching results against the 80% bar, measure by measure, from stored benchmark runs (superuser; dev only)

#How to Connect

#The MCP Server URL

https://mcp.ambientchat.ai

That's the address. Point any MCP-compatible client at it, sign in with your ambientChat credentials, and the tools above become available.

#Loading only some tools (toolsets)

By default an MCP client loads every tool above — over 200 of them, around 60,000 tokens of the AI's attention before you ask anything. Claude Desktop, claude.ai and Claude Code load tools on demand, so they don't need this. A plain agent that loads everything on connect can ask for less by naming toolsets in the URL:

https://mcp.ambientchat.ai/mcp?toolsets=inventory,documents

(Or send the header X-MCP-Toolsets: inventory,documents.) The toolsets are inventory, documents, views, sharing, conversations, location, settings, account, knowledge, data, context and diagnostics. A small core always comes along: your context, help, journaling, skills, query_expert, and the few tools those hand off to (reading a web page, checking an export, a skill job or an expert's answer). The agent is told at connect which toolsets it left out, so it can say so rather than guess. To get everything back, connect without toolsets.

#Via Claude Desktop (MCP)

  1. Open Claude Desktop settings
  2. Navigate to the MCP Servers configuration
  3. Add a new MCP server with the URL https://mcp.ambientchat.ai
  4. Authenticate with your ambientChat credentials
  5. Claude Desktop will discover all available tools automatically

#Via Meta's Muse (or any app that speaks REST, not MCP)

Muse can't plug into an MCP server, so ambientChat offers the same tools as a web API it can call. Step-by-step setup, what Muse can see and how to pause or disconnect it: Connect Muse — Use ambientChat from Meta's Muse.

#Using Tools in Conversation

Once connected, you don't need to call tools by name. Just ask natural questions and the AI uses the appropriate tools:

  • "What food is expiring this week?" → list_expiring_items
  • "Find my Costco receipts from January." → search_documents
  • "How much is my electronics collection worth?" → search_items
  • "Create a packing list for my Hawaii trip." → create_markdown_document
  • "Change all items in the kitchen to category Food." → update_items_by_filter
  • "Undo that last bulk change." → undo_bulk_change
  • "Export all my data as CSV." → export_user_data
  • "What did we talk about last week?" → recall_history
  • "Share this recipe with my wife." → share_conversation

#Tips & Tricks

  • Same context everywhere. MCP gives external clients the same context as in-app chat. The search quality, semantic matching, and data access are identical.
  • Great for desktop workflows. Writing a report? Planning a meal? Doing insurance paperwork? Your full inventory and document archive is accessible without switching to your phone.
  • Combine tools naturally. Search your inventory, then create a document summarizing what you found. The tools compose together — search, analyze, write, save — all in one conversation.
  • Everything is undoable. Bulk changes, individual edits, trash operations — all reversible. Say "undo that" and the AI knows what you mean.
  • Voice works too. Every tool listed here works identically via voice in the mobile app. Just say what you need.

#Known Limitations

  • Requires a compatible MCP client — or a REST-capable agent. Claude Desktop is the primary supported client. The MCP server follows the standard protocol; apps without MCP (Muse) use the REST API at /v1, which offers the same tools.
  • Uploading images/files now works via MCP. Use upload_document with inline base64 (small files) or the get_document_upload_url → PUT → upload_document flow (any size). In claude.ai chat, dropping in a photo and saying "save this to my Medical folder" works when code execution is enabled — the assistant uploads the attached file for you. Live, on-device camera scanning still happens in the mobile app.
  • Rate limits apply. MCP tool calls consume the same credits as in-app chat. Your subscription tier determines available credits.
  • Real-time sync is not instant. If you scan something on your phone, it may take a few seconds before MCP tools see the new item.

Staff account operations use the same list_users_directory contract with an explicit operations request, account audience, sort and direction. The server orders the complete authorized operational snapshot before pagination. Publication-only staff receive no operational account fields.

#Version History

Version Date What Changed
191 2026-11-17 When you ask how to stop or change one of your connected AI apps, the chat now opens Context Sharing with that app's row circled, not just the screen. AMB-3212
190 2026-11-17 The coach's screen list now has room for new screens: it lists every screen by area, and gives the full description of each only when a request names one it does not know AMB-3212
189 2026-11-17 Asking the coach about a folder from inside My AI Context now circles that folder's Share, not the one you were looking at AMB-3212
188 2026-11-17 Ask the coach how to share a folder, and Take me there opens that folder with its Share circled, in the app and on the web AMB-3212
187 2026-11-17 Ask the coach where you have been and it opens the Location tab with Journey history circled; while Location history is off, it offers to turn it on in Control instead AMB-3212
186 2026-11-17 The coach no longer offers to take an account that signs in with Apple or Google to change a password it does not have, and on the phone it circles Find in document in the ⋯ menu AMB-3212
185 2026-11-16 AMB-3321: Inventory search finds empty and nested folders by name on iOS, web, chat and MCP, with full paths and separate item counts.
184 2026-11-16 Document the read-only, capability-scoped list_users_directory tool (AMB-3319)
183 2026-11-16 On the phone, Take me there now scrolls the circled switch into view when it is further down the screen, and finds a beacon far down your list AMB-3285
182 2026-11-16 Ask how to change the voice or how fast it speaks, and Take me there opens Voice with that choice circled AMB-3212
181 2026-11-16 When Claude or ChatGPT sends a screen to your phone or browser to help you change one setting, Open now takes you there with that control circled, as the Take me there button in ambientChat does AMB-3212
180 2026-11-15 Document search tools return matching owned folders with canonical paths alongside documents, including empty and nested folders (AMB-3304).
179 2026-11-15 Folder discovery, create/move results and save receipts identify the full Home destination; assistants distinguish same-named folders by path and real ID. AMB-3303
178 2026-11-15 Added additive website roles, scoped publishing and analytics access in the existing console; preserved the canonical help and deferred newsletter-delivery workflows. (AMB-3302)
177 2026-11-14 Ask how to turn off sharing with one AI app, model, beacon or muted venue by its name, and Take me there circles that one's switch, or the whole list if it isn't there AMB-3283
176 2026-11-13 show_screen can open one of your own things: Take me there for "how do I get to my Music folder?" opens that folder, titled with its name, on the phone and the web; an outside AI gets the folder's own link. "Where's my Music folder?" still gets the links and cards in the answer, and no button AMB-3212
175 2026-11-13 When an outside AI such as Claude or ChatGPT points you to one of your own things, it now arrives as a card on your phone or the browser you are using, as a screen does, and its link opens the ambientChat app on that thing AMB-3212
174 2026-11-12 list_conversations and recall_history say when an answer stated world facts from memory with no web check (context_receipt.web_check, with the reason); for an answer someone shared, recall_history and get_shared_resource say it as web_unchecked; the account export's conversations.csv has a receipt_web_check column AMB-3255
173 2026-11-12 show_screen can circle one control: Take me there for "how do I turn off access alerts?" opens Context Sharing with Notify when accessed circled. A control inside a row or section that must be opened first gets that circled first, unless it is already open; a control it cannot find on that screen just means nothing is circled, and the screen still opens AMB-3212
172 2026-11-11 show_screen opens the section an answer is about: Take me there for "which AIs can access my context" now opens Control at Context Sharing, not the top of Control — and every page in Control's search (Profile, Security, Billing history, Plans, Voice, Taste shares, Location precision, Places, Walk & Teach, Ranging, Application behavior, Don't Panic, Delete account) can now be a Take me there destination AMB-3205
171 2026-11-11 share_taste takes app_keys — several connected AI apps in one call, all or none — and replies with grants (AMB-3193)
170 2026-11-09 get_taste rows carry brand (normalised; a passport's issuer first) and a new clothing kind. AMB-3141
169 2026-11-08 get_taste returns the taste summary; update_user_setting profile.taste_summary edits or regenerates it. AMB-3071
168 2026-11-03 New verify_item_value tool: a sourced web price for an item valued at $250 or more. AMB-2869
167 2026-10-29 Disconnecting an AI app in Context Sharing now also cancels any sign-in it had started but not finished, so it cannot sign straight back in AMB-2808
166 2026-10-29 Disabling or deleting an ambientChat account now signs it out of every connected AI app within a minute, including apps that were mid-session, and a deleted account can no longer be brought back by an app that still held its access AMB-2807
165 2026-10-29 A disabled or deleted ambientChat account can no longer sign in to, or stay signed in to, a connected AI app: the app now shows a sign-in error instead of a connection that fails on every request AMB-2806
164 2026-10-29 When a connected AI app's expert answer is cut off mid-stream it now gets the answer (or its ticket) instead of being told to ask again and pay twice; a disabled account is told so rather than retried; and a rate limit, including web search's, says to wait and for how long, where it used to read as a generic failure AMB-2805
163 2026-10-29 When a call from a connected AI app (Claude, ChatGPT and others) or the developer API fails between our servers, the app now tells it in plain words, and says when a change may already have been saved, never our server's internal error text AMB-2801
162 2026-10-29 list_markdown_documents takes a cursor and returns next_cursor when truncated, so an agent told there are more notes can fetch the rest, page by page, with a tag too AMB-2792
161 2026-10-29 list_markdown_documents says truncated: true when more notes matched than limit, so count is never read as the total; search_markdown_content word fallback (when meaning search finds nothing) now looks through every note, not the first 200 AMB-2792
160 2026-10-29 list_markdown_documents with tag finds every tagged note, however many notes there are; it matched the tag only among the newest 250 before AMB-2782
159 2026-10-28 list_document_versions, get_document_version and recent_list_changes give every row an actor_label, who made it in the words to show (the ones recent_item_changes uses); recent_list_changes also gives its actor_kind AMB-2757
158 2026-10-28 update_shared_resource on an item or document its owner has deleted now refuses with resource_deleted ("Steve deleted this…"), as opening it does, not "Item not found" AMB-2731
157 2026-10-28 Tool permissions now have two settings per tool, ambientChat chat and Outside AIs, on the phone and the web. A tool you block for outside AIs is refused to ChatGPT, Claude and other AI apps; they can never see or change your permissions AMB-2645
156 2026-10-28 Ask ambientChat's chat to set tool permissions for you — "block anything that deletes for outside AIs" — and it shows the changes as a card; nothing changes until you tap Apply (propose_tool_permissions) AMB-2645
155 2026-10-27 New tool coach_steps: the AI's how-to steps are shown on the real controls of the screen, one at a time, while you do the task AMB-2585
154 2026-10-27 New tool show_screen: points to the app screen where a how-to answer's steps happen — a Take me there button in ambientChat, a web link or a "do this on your phone" sentence for outside AIs AMB-2584
153 2026-10-27 The screen-request card can now appear in your web browser too, when you are using ambientChat there and the web has that screen. Phone-only screens still go to your phone, and get_coach_status callers see where it went in sent_to AMB-2584
152 2026-10-27 An outside AI's link from show_screen or coach_steps now opens the ambientChat app at that screen on a phone that has it, and the web page anywhere else. Phone-only screens get a link too AMB-2584
151 2026-10-27 Adds propose_field_values, which fills in a form for you to check and save. It writes nothing; other AI apps are pointed to the regular, undoable tools AMB-2584
150 2026-10-27 Adds get_coach_status: an outside AI that sent a screen to your phone can check whether you opened it and how far you got AMB-2584
149 2026-10-27 Shows the phone card an outside AI's screen request arrives as, with Open and Not now. There is no web card yet AMB-2584
148 2026-10-25 get_help returns the section that answers. It used to return the one paragraph that matched most of your question's words; it now returns the whole section of the guide that answers it — at most about 500 words — with the guide's title, the section's heading and a link straight to it, and short pointers to other nearby sections AMB-2526
147 2026-10-25 Agents connected to ambientChat know what it does. Claude, ChatGPT, Muse and any other agent now receive a short map of ambientChat's features and where they run when they connect, and are told to look up the help with get_help before answering a "how do I…" question about the app, instead of guessing at its steps AMB-2526
146 2026-10-22 Added the superuser-only get_tag_location_rollout (who has server-side tag location, in shadow or on, and who last changed it) and set_tag_location_rollout (move accounts in or out by uid or email; putting anyone on needs confirm, and mode off rolls back) AMB-2393
145 2026-10-21 list_image_models now says how many reference photos one image uses (max_references), and generate_image reports any it could not use (references_unused, with a note for you) instead of leaving them out without a word (AMB-2199)
144 2026-10-15 An MCP client can load only some tools: connect to /mcp?toolsets=inventory,documents (or send X-MCP-Toolsets) to get the always-on core plus those toolsets, and the agent is told which ones it left out. Connecting without toolsets still loads every tool (AMB-2063)
143 2026-10-13 Added send_test_access_alert — a test of the AI-access alert to every registered device, with Apple's (or Google's) answer for each (AMB-2056)
142 2026-10-13 Added the superuser-only get_home_build_report (each Desktop AI build's speed and failure rate over the last days, official builds apart from local experiments); get_home_fleet_versions now says where each Mac's build came from (official, local, or an uncommitted experiment) AMB-2009
141 2026-10-13 New recent_list_changes and undo_list_change: see who changed a checklist and through which AI app, and undo a change by voice ("undo that"). A removed line comes back with its tick and its place, and an undo of a line somebody has changed since is refused (SUPERSEDED). Rows now carry added_via / checked_via, and every row write names the list and the line AMB-1991
140 2026-10-13 The list row tools work on a list someone shared with you — pass owner_uid; rows say who added and ticked them by name, and the read says whether you can edit AMB-1988
139 2026-10-07 list_available_models says Desktop AI is free: your Mac's entry carries cost_multiplier 0, beside the $×N of every cloud model, so an assistant asked which model is cheapest names your Mac AMB-1889
138 2026-10-07 Added remove_home_node — remove one of your Macs from Desktop AI for good; its credential stops working and it leaves your model settings (AMB-1887)
137 2026-10-07 Added the superuser-only get_home_fleet_versions (which Desktop AI version each Mac runs, who is behind, who rolled back) and set_home_agent_rollout (founder → canary → all, gated on speed and failure rate) AMB-1796
136 2026-11-16 get_context_sharing now returns every connected AI app with its own switch (apps), and its master switch reads as enforced: an account that never set it is sharing. The retired per-service flags (Claude, McCormick, Amazon) are gone from it, so an AI asked "is Grok sharing my context" no longer answers from a list Grok was never in (AMB-3284)
135 2026-11-13 generate_image / list_image_models: GPT Image 1 Mini and GPT Image 1.5 are gone, because OpenAI shuts them off on 1 December 2026. GPT Image 2.5 Flare · Low (gpt-image-2.5-flare-low) takes Mini's place as the cheap, fast OpenAI option and the OpenAI draft default, at about the same price. Product pictures, folder icons and profile-photo Reimagine draw with GPT Image 2.5 Flare too
134 2026-11-11 apply_receipt_price (new): an item's offered receipt price becomes its Price. link_document_to_item on a receipt sets the item's Price to its line and says so in receipt_price; get_item carries price_incl_tax, receipt_store, receipt_date and receipt_price_proposal (AMB-3195)
133 2026-11-11 query_expert takes web_search (default false). When it is true, the server searches the web once on the question and sends the expert the results, for any model. Otherwise no expert searches, Claude included, which used to search on its own. The answer reports web_search (AMB-3171)
132 2026-11-11 A running get_expert_answer says how far along the expert is: elapsed_s, and stage (starting, thinking, searching, using_tools or writing). poll_after_s grows from 10 s to 30 s as the run gets longer (AMB-3172)
131 2026-11-11 get_expert_answer and query_expert say why an expert failed: failure_reason, stop_reason, partial_answer (incomplete) and suggested_fallback_model (a suggestion; never switched to automatically). A refusal, or a notice that the model used up its answer budget, is now a failure, not a completed answer (AMB-3172)
130 2026-11-10 improve_item_photos takes mode: "beautify": an AI-drawn studio product shot of the item's camera photo, display only (stored in styled_image_urls, improved_review.kind: "styled"), priced at the image rate and charged only for a delivered picture (AMB-3045)
127 2026-11-10 Added set_item_edition — pick which record an album is from get_dpp → music.match, or confirm the one matched; undoable with undo_item_change; gather: true starts the one lookup for other editions. music.match carries settled, switching and gathering; get_dpp only reads (AMB-3034, AMB-3164)
129 2026-11-10 Added revisit_item_match — Take a closer look: a paid AI comparison of the user's photo with the candidate sleeves recommends a record (never picks it). music.match carries can_revisit, revisit_credits and revisit; accepting it through set_item_edition records decided_by: "revisit" (AMB-3036)
128 2026-11-10 set_item_edition takes hint (a catalogue number or barcode): one free MusicBrainz search adds a candidate; music.match.hint says what it found (AMB-3035)
126 2026-11-10 list_items rows carry match_ambiguous (a possible mix-up); page 1 carries ambiguous_count, and only_ambiguous lists just those albums. get_dpp now shows music.match for albums matched before the journal existed (AMB-3034)
125 2026-11-09 Added improve_item_photos — Clean up: re-frame and re-light an item's own photo, priced first (no confirm), then run (confirm: true). The cleaned picture becomes the pinned default, the original is kept; get_item lists it in cleaned_image_urls (AMB-3043)
124 2026-11-08 finish_item takes add_to_list (and list_id to name the list; an id that is not one of the user's lists still finishes the item, adds no line, and says why in list_row_error) and answers list_row (row_id, list_id, list_title, already_on_list): one line on the shopping list linked to the item, never two — say which list. Undoing that finish (unfinish_item, or undo_item_change on its event_id) takes the line back while it is still unticked and unedited (list_rows_removed). Ticking a line linked to a Finished item now brings that same item back, with today as its purchase date: set_list_row_checked answers item_restored (set_list_rows_checked: items_restored). On a list someone shared, a line the owner's own finish put there comes back for anyone who may tick it (recorded in the item's history as them); any other linked line only for someone who may edit the item — otherwise item_unchanged. Unticking never finishes it again. undo_list_change on that tick finishes it again too (items_refinished). add_list_row's description now says exactly this (AMB-3070)
123 2026-11-08 finish_item takes finished_sale_price (sold only) and answers warranty_documents (broke) and follow_up, the follow-ups the apps' finish sheet offers. Items carry finished_sale_price. list_trash rows carry status, so a finished item says so before it is restored (AMB-3069). On an item someone shared with you, finished_reason and finished_sale_price are left out: you see that it is finished and when, not why (AMB-3092)
122 2026-11-08 Added get_taste — your 👍 / 👎 on items and documents (finished items included, each with its kind) and your interests, from the same reader chat uses (AMB-3068)
121 2026-11-08 Added finish_item and unfinish_item. Items carry status, finished_reason and finished_at; list_items leaves finished items out, reports finished_count and takes include_finished; export_folder takes include_finished; create_items_bulk brings a finished match back (brought_back) instead of skipping it (AMB-3067)
120 2026-11-07 Added set_reaction (👍 / 👎 on an item or document). Items and documents carry reaction and reacted_at; list_items and list_documents take reaction_first. The never-written like_score and liked_at are gone from both (AMB-3066)
119 2026-11-05 start_home_node_setup keeps a Mac's model when it is set up again, picks one for the Mac's memory otherwise (optional ram_gb, from the Desktop AI app, for a Mac that hasn't reported yet), and says which and why in model_build_label, model_build_reason and model_build_note (AMB-2882, AMB-2967)
118 2026-10-29 get_request_receipt: whether an item or document edit sent with an action key ran, read from its request receipt — landed, not_applied, running or unknown; only landed means it took effect. It is for the phone's outbox, which does not call it yet (AMB-2820)
117 2026-10-28 undo_bulk_change: each entry in skipped now carries a code for why it was left alone (STALE_EVENT, NOTHING_TO_UNDO, ALREADY_UNDONE, NOT_FOUND, PICTURE_GONE, NOT_UNDOABLE), the same codes undo_item_change refuses with; reason stays (AMB-2661)
116 2026-10-28 recent_item_changes gives every change can_undo: whether undo_item_change would put it back now, decided by the same rule (false once undone, changed again since, or a create, trash, restore or sighting). The app's Recent changes cards offer Undo exactly there (AMB-2638)
115 2026-10-28 recent_item_changes gives every change an actor_label, who made it in the words to show: "You", "You, via chat", ", for you", "Shared with you: " or "ambientChat" (AMB-2627)
114 2026-10-28 import_user_data still takes no tag back; the summary and the dry run gain tags: how many are already the user's, and every other tag in tags.unresolved with reason and message, kept as a tag_bindings row that the next backup carries until the user claims the tag. release_tag now records who released it (AMB-2404)
113 2026-10-28 import_user_data restores the places and Zones in a backup as your own (new, private, each keeping its source id as logical_id); a second import finds them rather than copying them. The summary and the dry run gain places. Beacons and tags are still not restored (AMB-2404)
112 2026-10-28 Added release_tag — the owner gives a tag up so another account can set it up; it also leaves its Zone, its links and tracking are removed, and its name, notes, alert text and group are cleared (AMB-2404, AMB-2704)
111 2026-10-27 query_expert answers a retry with the same _meta.idempotency_key once: from the first answer with replayed: true and credits_used: 0, or the first call's ticket while it runs. The same key with another question is refused REQUEST_ID_REUSED. rescan_document and upload_document likewise queue one scan per key (AMB-2674)
110 2026-10-27 recent_item_changes: an edit someone you share an item or document with makes through update_shared_resource now reads changed_by: 'user:{their uid}', naming them, not you; the source and their AI app are kept. A picture they rotate, upload or set as the face does the same. recent_item_changes' output schema says what user:{uid} and link:{id} mean (AMB-2644)
109 2026-10-27 recent_item_changes: an edit, bulk edit or bulk add made by the in-app chat now reads source: 'ai_chat', and names you and the app in source_details.edited_by (an outside AI app's reads source: 'mcp' with its name there), the same as a text save from either (AMB-2627)
108 2026-10-27 undo_item_change and undo_bulk_change record the undo as whoever really made it — you, the in-app chat, or the AI app working for you — never an actor the caller names: their source and changed_by arguments are gone, and are ignored if still sent (AMB-2627)
107 2026-10-26 Added share_list_link — make a list's public link that whoever holds it can tick, with no account, or switch an existing one to look-only. get_share_status link entries now say can_tick (AMB-1486)
106 2026-10-25 link_beacon gives a tag the one-item rule whoever links it: staff linking an item onto someone else's tag now replace its item (it used to add a second one), keep its documents, and the item shows which tag is on it
105 2026-10-24 query_expert include_documents also takes a document someone shared with you, sent in full and credited to its owner; a share that has ended is reported in include_documents_unconfirmed (AMB-2332)
104 2026-10-23 share_with_email takes your own Desktop AI Mac (home_node): the person takes one of its 5 seats, or is seated when they join if they have no account yet; a sixth is refused with home_node_seats_full (AMB-2375)
103 2026-10-23 remove_interest removes interests only: given the id of anything else, such as one of your settings records, it refuses and deletes nothing (AMB-2408)
102 2026-10-23 share_with_group keeps the length you choose (expires_in_days), on the phone, the web and for AI agents; it was always 30 days (AMB-2394)
101 2026-10-23 Desktop AI sharing: share_with_group takes a Mac (home_node); list_home_node_people, approve_home_node_seat, remove_home_node_seat and get_desktop_ai_guest_usage for the owner; answer_home_node_default_offer and acknowledge_home_node_disclosure for a guest; list_available_models marks a shared Mac with shared_by (AMB-2375)
100 2026-10-23 list_user_settings now includes your interests (what the AI knows you're into), every one of them. They still change with add_interest and remove_interest (AMB-2404)
99 2026-10-23 list_user_settings now includes your tool permissions (the actions you always allow or block) and your connected AI apps (whether each may read your context, and when you gave or withdrew that). list_setting_dimensions names the tool that changes each: update_tool_permission and set_connected_app_sharing, not update_user_setting, so their safety checks always apply — an outside AI app still can't switch one back on (AMB-2404)
98 2026-10-22 claim_tag is withdrawn as a tool: making a tag yours needs the phone's camera on its label and its radio hearing it, and an agent can only make both up — for an Identiv tag, both are the number printed on the label. Set a tag up by scanning it in the phone app. Agents are no longer offered it; one still holding the old tool list is told to use the phone (CLAIM_NEEDS_PHONE) and nothing is written (AMB-2319)
97 2026-10-22 file_items and move_items_to_view return an op_id (and a display_summary) whenever an item actually moved; pass it to undo_bulk_change to put every item back in the folders it was in. Filings also show in recent_bulk_changes (AMB-1446)
96 2026-10-21 query_expert reviews a whole document by id: include_documents takes any of your documents, sent in full, and naming documents leaves your personal context and tools out (context: "none", or "auto" to keep them). A long answer no longer dies at your agent's tool timeout — after about 40 s the call returns a job_id, and the new get_expert_answer collects the answer (AMB-2317)
95 2026-10-20 set_private_ai_access switches Desktop AI on for any account in the main app too — no invitation; private_ai_access is never unavailable any more (AMB-2234)
94 2026-10-19 list_document_versions, get_document_version and save_document_version_as_copy take owner_uid: someone a note is shared with can see and copy its saved versions; the copy lands in their own documents (AMB-2046)
93 2026-10-19 Added list_image_models — each image model's cost as a $×N and in credits, so an outside AI can say what an image will cost before it picks a model (AMB-2014)
92 2026-10-18 Added email_desktop_ai_setup_link — emails you the Desktop AI setup link to open on your Mac (AMB-2133)
91 2026-10-18 set_private_ai_access with no email switches Desktop AI on or off for your own account; list_available_models and get_ai_model_pref report private_ai_access (superuser, member, available, unavailable) (AMB-2133)
91 2026-10-16 The Settings tab is now called Control, on the phone and the web; the web address is /control (old /settings links still work) (AMB-2120)
90 2026-10-16 An outside AI can tick the line it means on a long list. set_list_row_checked takes text ("string cheese") instead of row_id, and set_list_rows_checked takes texts: an exact line wins, two possible lines are refused with both named (ROW_TEXT_AMBIGUOUS), and a name on no line is refused with the closest ones (ROW_TEXT_NOT_FOUND) — nothing is guessed. list_list_rows takes query and unchecked_only, and its rows carry only what is set (an 83-line list now reads back in under 10 KB). A row id that is not on the list says where row ids come from (AMB-2096)
88 2026-10-13 The list row tools (list_list_rows, add_list_row, set_list_row_checked, set_list_rows_checked, update_list_row, move_list_row, remove_list_row) take owner_uid for a list someone shared with the user: reading needs view access, any change needs edit access (READ_ONLY otherwise), and private: true ("just me") works only on a line the caller added (NOT_YOURS otherwise). Rows now carry added_by_name / checked_by_name, and the list read says the caller's access (AMB-1988)
89 2026-10-13 Added list_document_versions, get_document_version and save_document_version_as_copy — browse a note's saved versions (V number, when, who, and which AI app saved it), read one, and save one as its own editable document. A whole-note rewrite now also requires expected_source_hash (SOURCE_HASH_REQUIRED without it) — AMB-1972, AMB-2036
87 2026-10-13 find_album_covers reports albums whose lookup was already waiting as already_pending, apart from queued (the ones it started). Asking again during an outage no longer claims to have started lookups it didn't (AMB-2018)
86 2026-10-09 Listed check_scan_affordability — the low-balance check the app runs before a photo scan, so an agent can warn before a scan the user can't afford (AMB-1906)
85 2026-10-09 Listed present_records. Its cards now say where each item and document is filed ("in Garage", "at Home"), in the structured folders/at_root fields and in the markdown block Claude Desktop and other apps show (AMB-1815)
84 2026-10-08 update_beacon takes beacon_name — rename a beacon or tag (a tag's new name reaches the item it tracks). New delete_beacon deletes a beacon or tag after you confirm; a deleted tag's item stops being tracked (AMB-1741)
83 2026-10-07 link_beacon takes unlink: true to remove ONE link by kind + id and keep the rest (on a tag, unlinking its item untracks it). list_beacons and list_asset_tags report each record's linked_refs, plus missing_linked_refs for links whose item or document was deleted (AMB-1740)
82 2026-10-07 Added trash_item_image: move one of an item's pictures to Trash; restore_trash puts it back
81 2026-10-07 update_beacon takes context_string — the words handed to the AI whenever the beacon or tag is detected (empty string clears it). list_beacons and list_asset_tags report it, as plain words
80 2026-10-07 Added upload_item_image (the user's own picture becomes the item's thumbnail) and set_item_face_image (choose or pin which picture is the thumbnail); items carry original_image_url and face_pinned_url
79 2026-10-06 start_home_node_setup takes model_build_id (superusers); setting a Mac up again moves it to the current Desktop AI model, Qwen 3.6 35B (AMB-1829)
78 2026-10-05 Private AI is now called Home AI in every tool description and result: a Mac is listed as “your Mac's name · Home AI”. Tool and field names are unchanged (set_private_ai_access, private_ai_enabled) (AMB-1782)
77 2026-10-05 Added get_beacon_snoozes, and update_user_setting takes key beacon_snoozes: snoozing a beacon is kept on your account, so it lasts as long as you asked, follows you to every device, and never silences the beacon for anyone else (AMB-1629)
76 2026-10-05 find_album_covers matches the album, not a single of the same title. It also re-matches albums that were given a single's sleeve and two-track listing, and counts them in queued (AMB-1771)
75 2026-10-04 get_dpp returns music for albums: the track listing, edition, genres, credits, notes and Discogs collectors' numbers, merged from MusicBrainz and Discogs. Item rows carry album_summary. find_album_covers also fetches track listings (AMB-1770)
74 2026-10-03 Added start_home_node_setup — set up Private AI on your own Mac with one Terminal command — and the superuser-only set_private_ai_access. list_available_models now says whether Private AI is on for your account (private_ai_enabled), and a Mac still being set up reads “Waiting for setup on your Mac” (AMB-1752)
73 2026-10-03 Share durations are now chosen as '30', '60', '90' or 'never' (strings) on share_with_email, update_share_grant and renew_share_grant, where they used to be a number or null. update_share_grant's expires_at takes a number only — to make a share never expire, pass expires_in_days: 'never'. update_beacon clears alert text with an empty string, and update_list_row / add_list_row unlink an item with an empty item_id, instead of null. Nothing you can do changed; the old null form broke Gemini (AMB-1700)
72 2026-10-02 Added renew_share_grant — renewing an in-app share never shortens it. update_share_grant no longer claims "give them longer", reports the expiry it actually wrote (it used to echo the old one), and returns previous_access / previous_expires_at for undo (AMB-1540, AMB-1497)
71 2026-10-01 Added list_available_models — the models you can pick, including your own Mac's Private AI with its online state — and the superuser-only register_home_node. get_ai_model_pref and update_user_setting (key ai_model) accept a Private AI as your default, and report it unavailable with the reason while the Mac is offline (AMB-1644)
70 2026-09-30 Added update_beacon — change a beacon's alert mode, alert text, prominence or live/paused on its registry record, as a partial update. list_beacons and list_asset_tags now report alert_mode, alert_string and prominence, so an agent can read a setting before changing it (AMB-1444)
69 2026-10-01 Muse setup moved to its own guide, Connect Muse
68 2026-09-30 REST agents (Muse) can now read the same connect-time guidance as MCP clients — your skills list, the journaling policy and the skill-authoring rule — at GET /v1/instructions. A paused or switched-off app no longer learns your skills list when it connects, over MCP or REST (AMB-1616)
67 2026-09-30 Added find_album_covers: real album cover art in place of AI illustrations; list rows now carry image_regenerating (picture on its way) and cover_art_source (AMB-1621)
66 2026-09-30 Added list_connected_apps, set_connected_app_sharing and disconnect_connected_app: every connected AI app gets its own switch, checked on every request over MCP and the REST API (AMB-1615)
65 2026-09-30 query_expert now obeys your context-sharing switch and appears in your access log like every other tool — with sharing off, no connected AI can reach your context through the expert panel (AMB-1614)
64 2026-09-30 Every tool is now also a REST API at https://mcp.ambientchat.ai/v1 (OpenAPI at /v1/openapi.json), same OAuth sign-in — so agents without MCP, starting with Meta's Muse, can connect (AMB-1613)
63 2026-09-28 create_journey_report refusals: an empty folder is now code no_items (was invalid_request), and too_many_items carries details: { asset_count, max } (AMB-1513)
62 2026-09-28 create_journey_report takes up to 25 asset_ids or a view_id (a folder, followed live); the page draws several items as lanes on one clock with a temperature line per item. get_asset_journey's temperature points now carry asset_id (AMB-1172)
61 2026-09-28 Added create_journey_report — publishes a tracked item's chain of custody as a live, no-account page through the normal document public link; asking again with the same settings returns the same link (AMB-1172)
60 2026-09-27 Added list_import_jobs — your recent imports, so a restore can be checked on after leaving the screen that started it (AMB-1366)
59 2026-09-25 Added get_asset_journey — a chain of custody for one or more tracked items over a required date range: chapters by venue, stops by zone, dwell times, temperature, and the underlying events. It says when it is partial instead of implying completeness (events_complete, temperature_complete, assets_complete). list_item_location_history gains optional from/to, and list_item_temperature_history now reports complete (AMB-1171)
58 2026-09-25 file_items and update_items_by_filter take in_any_folder: true — every item filed in some folder, the opposite of at_root (passing both is refused). "Move everything out of every folder" is one short call instead of an agent listing and re-emitting every filed item's id (AMB-1346)
57 2026-09-23 get_account_deletion_precheck reports your latest export (latestExport) and whether one is still being made (exportInProgress), so an assistant can offer — never require — that you save a copy before deleting (AMB-1360)
56 2026-09-22 The voice settings gain tour_listen, the Radar's Zone guide (on by default): read by get_voice_preferences, changed with update_user_setting key voice
55 2026-09-22 link_beacon refuses a target that does not exist (structured, naming it) and never carries a beacon's dead links forward; list_asset_tags reports linked: null for a tag whose item is gone rather than the ghost's name
54 2026-09-22 resolve_scanned_code is listed here for the first time, and now knows a Bluetooth tag's printed label: it used to answer Not recognized and echo the label as printed, while the phone called it a Tag. It now answers a Tag card with tag.address in the standard spelling whatever the label used, so an assistant can look the tag up with list_asset_tags
53 2026-09-21 get_ble_scan_rate and the ble_scanning setting are withdrawn: the Bluetooth scan rate was a Settings → Bluetooth choice that page no longer offers — the phone tunes its own radio (see Beacons → How the phone listens). "Set my scan rate to eco" no longer does anything; there is nothing to set
52 2026-09-18 reserve_device_identity answers prior — what the row was before the call — so the phone reads a previously set-up device's current values and asks before overwriting them. POST /resolve (the phone's own lookup, not an agent tool) now says what a tag is to the caller: how far along its set-up is and whether it is yours, somebody else's or nobody's — never who
51 2026-09-18 confirm_device_provisioning is withdrawn as a tool: the report of a tag set-up is the phone's, over its own channel, in the same class as a tag sighting — an agent that could file “verified” would only be fabricating it. What an agent needs is on list_asset_tags, which now carries provisioning_state, firmware, oem_model and flashed_profile, and no longer lists a tag you have never had while it is being set up (or after a failed first attempt) — a tag you already had stays listed, with its state. reserve_device_identity checks the whole registry by MAC: a tag you already claimed is adopted (same row, now with an identity), one somebody else owns is refused before any plan is handed out, and asking on a different network returns the identity the tag already has
50 2026-09-18 reserve_device_identity now returns the write plan with the identity (profile, slots, interval, power, measured power) and chooses the network when none is named; the password candidates ride the phone's channel only, never an agent's. New confirm_device_provisioning — the phone's report flips a reserved row to active or failed, records what was intended and what the tag actually holds, and gives the device the same ownership shape a claimed tag has
49 2026-09-18 file_items joins the folder tools, and files by filter — a category, everything in Home, everything in another folder — so an agent never lists and re-emits hundreds of ids; update_items_by_filter takes the same folder filters (AMB-1203)
49 2026-09-19 get_scan_status’s capture_gate now says WHY an “only capture…” instruction was not applied. An agent asked “what happened to my scan?” could previously only report that the instruction failed, so it would suggest trying again — when the real answer was that the photo held a single subject and there was nothing to filter
48 2026-09-18 Added restore_skipped_objects, and get_scan_status now reports capture_gate — what an “only capture office furniture” scan kept, what it skipped and why. An agent could already be told a scan was “complete, 1 item created” while thirty-nine objects had been dropped, and had no way to mention it or undo it. Now it can say what went and put any of it back, into the same scan session and the same place
47 2026-09-17 query_expert now tells your agent which large documents the answer used and which were held back (with the token cost), and takes include_documents / exclude_documents so the agent can pay for one or leave one out on the next call. Before this an agent in Claude Code or Claude Desktop got an answer that quietly omitted a big document with no way to know
46 2026-09-08 Added reserve_device_identity — anyone signed in can now claim an iBeacon identity for a device they are setting up, and the number is HELD while they work. Allocating one used to need a staff account, and the old route only read the registry: two people provisioning at the same moment computed the same Minor and the loser's save failed. Asking twice with the same MAC returns the same identity, so a retry after a failed device write costs nothing
45 2026-09-08 Added read_web_page — the assistant can read what a scanned page says now, not just its title. Page content is treated as quoted data, never as instructions
44 2026-09-08 Added save_web_pointer — scan a web link and keep it as a document you can rename
43 2026-09-07 Added promote_item and discard_scan_session — AI Lookup keeps a capture only when you say so
28 2026-07-28 Spreadsheet → inventory import: propose_import_mapping reads the columns (remembering a layout you have approved before) and create_items_bulk creates the items as one operation you can undo in one go
29 2026-08-10 Text documents are editable: update_text_document replaces a note/.txt/.csv/.tsv, with undo_text_update / redo_text_update stepping the history — a spreadsheet undo restores the previous file exactly. Read the editable text with get_document({include_source: true})
30 2026-08-11 update_markdown_document, undo_markdown_update and redo_markdown_update are retired — markdown is a text document like any other, so update_text_document / undo_text_update / redo_text_update serve it too. The generic tool gained section updates (section_heading / append) so nothing was lost
33 2026-08-16 link_beacon ADDS a linked thing to a beacon (up to 8) instead of replacing what is there; kind none still clears them all
34 2026-08-20 Added get_venue_invitation_optouts / update_venue_invitation_optout — the per-venue "don't show this invitation again" choices, stored on your account so they survive reinstalls
36 2026-08-20 Settings writes converged on ONE tool: update_user_setting(key, value) replaces update_privacy_preferences, update_voice_preferences, update_notification_preferences, update_context_sharing, update_profile, update_save_preferences, update_background_trails_pref, update_app_behavior_prefs, update_ble_scan_rate, update_ai_model_pref, and update_venue_invitation_settings. Per-domain get_* read tools remain; list_setting_dimensions names each domain's setting_key
35 2026-08-20 Those two became get_venue_invitation_settings / update_venue_invitation_settings: one pair for the whole domain — the master switch moved into your account too, and un-muting a venue from anywhere (chat, web, phone) now reaches every device. Venue invitations also appear in list_user_settings / list_setting_dimensions
40 2026-09-02 An unknown tag battery now says WHY: list_asset_tags rows and get_item’s tracking carry the tag’s stored MAC and telemetry_join (ok / missing / malformed). A tag registered without a usable MAC can never report battery or temperature at all — the address is what matches the telemetry broadcast — so an assistant can now tell you to add it from the device label instead of guessing the tag is dead
39 2026-08-26 Added list_item_temperature_history — a tracked item's temperature trail from its Bluetooth tag's telemetry (recorded up to once a minute), newest first, with from/to windowing. Ask "what temperature is my tape measure at?" The same data draws the temperature graph in the app's TRACKING manage sheet
38 2026-08-26 Added list_item_location_history — where an item has BEEN, newest first: manual "Last seen" sets, the link-time stamp, and live tag sightings, with paging for long trails. Ask "where has my tape measure been?" The same history appears in the app when you tap a tracked item's TRACKING row
42 2026-09-05 You can create your own places: create_venue, update_venue and create_zone no longer need a staff role. A place you make is personal — only you can see it — because the alternative would have published your home address: the underlying registry treats a record with no explicit permissions as public, which is right for a museum beacon and wrong for a house. Two people can each have a place called "3M HQ" without either knowing the other exists; only a duplicate within your own places is refused. assign_beacons_to_zone stays staff-only — a beacon is shared hardware
41 2026-09-03 Added list_beacons — your FIXED beacons with battery state (good/low/replace/unknown), the reading's timestamp, last-heard, and the same telemetry_join explanation row 40 gave tags. A beacon's battery has always been recorded and was surfaced nowhere, which made a dead beacon indistinguishable from bad signal; list_beacons is the readable half. A beacon that reported once and then went silent is distinguishable from one never heard from, via battery_seen_at
37 2026-08-22 Asset tracking: link_beacon now lets you link a Bluetooth tag YOU OWN to your own item (no staff needed — the tag then tracks that one item and its category becomes Asset Tracker); new list_asset_tags lists your tags and what each tracks; items now carry a tracking fact (tag_id, tag_name) in get_item / search_items / list_items
27 2026-07-28 Spreadsheets: read_document_rows pages through a stored .csv/.tsv/.xlsx (one sheet at a time for a workbook, naming the others), and create_office_document is now create_data_file — it also makes .csv
26 2026-07-27 Word-style Undo/Redo — undo_markdown_update steps BACK one change (repeat to keep going), redo_markdown_update steps forward. A new edit discards anything ahead
23 2026-07-26 Added undo_markdown_update — document edits are snapshotted and can be reverted with "undo that"
22 2026-07-18 Added Tool Permissions section (list_tool_permissions / update_tool_permission) — per-tool Auto / Ask / Blocked policy, "Always Allow" on approval prompts
14 2026-07-05 Rewrote Sharing for in-app live sharing + groups (share_with_email, share_with_group, group lifecycle, revoke_share_grant, list_shared_with_me, …); added Connected Services section (add/list/update/remove_mcp_server) (v0.13.2)
13 2026-06-26 Added list_mounted_contents — resolve mounted shares into live data + provenance for inline display (AMB-377)
12 2026-06-26 Added mount_shared_resource / unmount_shared_resource / list_mounts — add shared content into your own library (AMB-376)
11 2026-06-25 Added list_skills / get_skill — first-party SKILL.md skills, readable by everyone (AMB-350)
10 2026-06-07 Terminology: undo surfaces say "the web interface" not "admin console"
1 2026-03-01 Initial guide — 28 tools across 6 categories
2 2026-05-27 Complete rewrite — all 93 user-facing tools across 17 categories (AMB-188)
3 2026-05-28 Added the MCP server URL (https://mcp.ambientchat.ai); reverted title to v1 to keep the seed idempotent
4 2026-05-29 Added get_document_upload_url; upload_document now accepts base64 / signed-URL bytes so MCP clients can upload images into folders (AMB-281)
5 2026-05-28 Added list_conversations (browse Q&A history with paging) — AMB-104 MCP promotion
6 2026-05-29 Added get_save_preferences / update_save_preferences (sticky save folder) — AMB-104 MCP promotion
7 2026-05-29 Added list_interests / add_interest / remove_interest (How I Roll tags) — AMB-104 MCP promotion
8 2026-06-01 get_profile / update_profile now read & write bio (folded in from the legacy profile field, which was removed) — AMB-104
9 2026-06-01 Added get_scan_status — query an upload/scan job by jobId (processing / complete / failed) — AMB-287