ambientChat lets you ask models from all three frontier labs (OpenAI, Anthropic, Google). You can put more than one of them on the same question, and you don't need a special mode for it. There are two kinds of second opinion, and they answer different needs:
- A review: switch model and ask. The new model reads the whole conversation, including the earlier answer, and knows which model wrote it. Use this when you want something checked, critiqued or improved: "Review GPT's draft and tell me what you'd change."
- An independent answer: ask an expert on its own. The other model answers without seeing the earlier answer, so it can't be swayed by it. Use this when you want an objective view, such as a second diagnosis, a valuation or a yes/no call, where you don't want the second model simply agreeing with the first.
Both work from coding tools connected over MCP too (Claude Code, Claude Desktop and others). See Via MCP below.
A second opinion doesn't have to come from another vendor. Asking a fast, cheap model and then that vendor's most capable one shows you whether the extra cost buys a better answer to this particular question.
#What It Does
Switching model mid-conversation keeps the whole thread. Once a conversation has answers from more than one model, each earlier answer reaches the next model tagged with the model that wrote it, e.g. [GPT-6.1 Sol]. So when you say "review GPT's answer", Claude knows exactly which answer that is and gives its own judgement rather than deferring to it. Nothing fans out automatically: you drive one model at a time.
Asking an expert on its own uses a tool called query_expert. Your current model sends a question to a different model, gets that model's answer back, and shows it to you labeled with the expert's name. If you ask, it compares the answer with its own. The expert never sees your conversation, only the question your current model sends it. That is what makes it independent. Ask for a review of one of your documents ("have GPT review my design doc") and the whole document goes with the question.
Why the difference matters. A model that has just read a confident answer tends to anchor on it. Asked for "a second opinion", it often polishes what's there rather than questioning the premise. In-thread review is great for catching mistakes and gaps. For a genuinely independent answer, the second model must not see the first one.
#How to Use It
#Get a review: switch model and ask (iOS and web)
What you need: nothing to set up. The picker lists the models switched on under Control → AI Model Selection → Enabled models. Every answer uses credits at that model's rate (the $×N badge in the picker). Your own Mac (Desktop AI) shows $0: its answers cost nothing, though a web search or image it makes still uses credits.
- Ask your question with whichever model is selected and read its answer
- Tap or click the model button: on the iPhone it's at the right-hand end of the row of round buttons above the message box; on the web it's at the top right of chat. The list, Choose your AI, groups the models by cost tier, with their capability icons. Pick a different model, from another lab or a different tier of the same one. Show me
- Either ask for what you want, e.g. "Review GPT's draft and tell me what you'd change" or "Do you agree with Claude's recommendation?", or leave the text box empty and press ↻. With nothing typed, the send button turns into ↻, which puts your last question to the model now selected
- Switch again for a third view, and so on. Each model sees the earlier answers and who wrote them
If it doesn't work:
- The model you want isn't in the list. It's switched off. Turn it back on under Control → AI Model Selection → Enabled models.
- A card says you've used your allowance. A second opinion uses credits like any question. The card says when your credits refill, and its button lets you continue now. Switching to your Mac doesn't get round a spent allowance.
- You picked your Mac and it's switched off, asleep or offline. Chat says your Mac can't answer right now and never switches model for you. Pick a cloud model, then press ↻ with the text box empty to ask again.
- "Couldn't get an answer. Please try again." Press ↻ with the text box empty to put the same question again.
#Get an independent answer: ask an expert on its own
Just ask in plain words, and say that the expert shouldn't see the earlier answer:
- "Ask Gemini, independently, for its own answer to my question. Send only the question and the facts it needs, not your answer or your conclusion."
- "Get an independent second opinion from Opus 5.5 on this valuation, then compare it with yours and tell me where you differ."
Your current model writes the question for the expert, so this wording matters. Ask it to include the facts the expert needs (the item, the numbers, the document's key terms), and to leave out its own answer. Then, if you like, ask it to compare the two answers.
The cleanest independent answer of all: a new conversation. Tap + (iOS) or New conversation (web), pick the other model, and ask the same question. That model starts with your personal context (your items, documents and profile) but nothing from the earlier thread. Bring its answer back to the first conversation if you want the two compared.
#Via MCP (Claude Code, Claude Desktop, and other AI agents)
The same query_expert tool is available to any MCP client you've connected to ambientChat, including Claude Code:
- Pass a
modelparameter with any currently-visible model ID (the tool description lists the available models and their capabilities) - To have it review a document, give it the document's id (
include_documents). Any of your documents is sent in full, whether or not it matches the question, and so is a document someone has shared with you (the expert is told whose it is), and the expert then sees only the question and those documents: none of your personal context and no tools, which is what a technical or blind review wants. Your agent can ask for your context as well withcontext: "auto" - Without named documents the tool includes your ambientChat context with each query: inventory, locations, and any of your documents that match the question
- A long answer is never lost. A reasoning model reading a long document can take minutes, longer than Claude Code and other agents wait for one tool call. If the answer isn't ready in about 40 seconds, your agent gets a
job_idback instead and collects the answer withget_expert_answer. Grok never collects ajob_idand gives up on a tool call after a minute, so for Grok the call waits for the answer itself for up to 50 seconds. If the answer takes longer, Grok asking the same question again collects it: for 10 to 20 minutes, the repeat is answered from the first run (waiting on it if it's still going), and isn't charged again. The answer is saved to your conversation history either way - A failed answer says why. If the expert can't answer, your agent is told the reason: the model declined, used its whole answer budget on thinking, took too long, its provider failed or was busy, or your credits ran out. Any words it had written come back marked incomplete, and your agent gets the name of another model from the same lab to try. That's a suggestion only. Your agent never switches models on its own, because you chose the model
- A large document is held back, not spent silently. Just as in the app (see Chat), a document big enough to crowd out the rest of your context is left out of the expert's answer and reported instead. The result's
documentslist names it withmode: "held_back"and the token cost of including it. Your agent can then ask you, or decide itself, and call again withinclude_documents: [id]to pay for it, or withexclude_documentsto drop one that was included. Each call decides for itself; nothing is remembered between calls
This is useful mid-work, not just for ambientChat questions. Reviewing a design document and want a second read before committing to it? Stuck on code that isn't behaving? Ask Claude Code to get a second opinion from a different model, e.g. "ask Opus 5.5 to review this function for edge cases" or "get Gemini's take on this design doc". It calls query_expert without you leaving your editor.
Example from Claude Desktop: "Ask Gemini what's the current market value for a 2023 MacBook Pro M3." The primary model calls query_expert to route the question to Gemini. If the answer needs real-time pricing, your agent asks with web_search on: the server searches the web once on the question and gives the expert the results first. Without it the expert doesn't search at all, whichever lab it comes from, so an independent opinion rests on the same footing every time.
For the full tool reference, see MCP Tools — Complete Reference.
#Tips & Tricks
- Draft with one model, review with another. For example, GPT drafts a contract, you switch to Claude for "give a second opinion on the draft, no changes yet", then you ask one of them to apply the changes you agreed.
- For high-stakes calls, get the independent answer first. Making a big purchase, checking health or legal information, valuing something? Get an independent answer before the models see each other, then ask one to compare. Two answers that agree independently mean more than one that agreed after reading the other.
- Watch for rubber-stamping. If the last model in a long thread says "this is ready, no changes needed", that may be the anchoring described above. Ask an expert on its own, or a fresh conversation, before you rely on it.
- Compare tiers, not just vendors. A fast model against a Premium one from the same vendor tells you whether the extra cost buys a meaningfully better answer to this question.
- ↻ works any time. With an empty text box, the send button re-asks your last question to the model now selected, on the phone and the web.
#Known Limitations
- No automatic fan-out. Your question is never sent to several models at once. You switch models and ask, one deliberate step per model.
- No side-by-side comparison view. Answers arrive one at a time as separate turns in the same thread.
- Independence depends on the question sent. When you ask an expert on its own, your current model writes the question. If you don't ask it to leave its own answer out, it may include it. Saying "independently" and "not your answer" is what keeps the expert blind.
- No automatic "best answer". The app never picks a winner among the models you asked. Ask one model to compare the answers, but the judgement is yours.
#Version History
| Version | Date | What Changed |
|---|---|---|
| 11 | 2026-10-27 | Getting a review now says what you need (the picker lists your enabled models; answers use credits at each model's rate, your Mac's cost nothing), where the model button is on the iPhone and on the web, and what to do when a model is missing, your allowance runs out, your Mac is offline or an answer fails AMB-2584 |
| 10 | 2026-11-11 | Via MCP: an expert searches the web only when asked (web_search: true). The server then searches once on the question and gives the expert the results, the same way for every model. Before, Claude experts searched on their own and the others never did (AMB-3171) |
| 9 | 2026-11-11 | Via MCP: a failed expert answer says why (failure_reason), returns any words written so far (partial_answer, marked incomplete) and suggests another model from the same lab (suggested_fallback_model), but never switches on its own. A refusal or an answer that used up its budget on thinking now reads as failed, not as a completed answer (AMB-3172) |
| 8 | 2026-10-24 | Via MCP and in the app: a document someone shared with you can be named for review by id too; the expert is told whose it is. A share that has ended can't be (AMB-2332) |
| 7 | 2026-10-21 | Via MCP: the expert can review a whole document by id (include_documents), seeing only the question and that document; and a long answer is never lost to the agent's tool timeout: it returns a job_id that get_expert_answer redeems (AMB-2317) |
| 6 | 2026-10-09 | The Panel of Experts toggle is gone, on iOS and web (AMB-1930). A second opinion is now simply: switch model and ask, or ask an expert on its own. Once a conversation mixes models, every earlier answer is tagged with the model that wrote it, so "review GPT's answer" works the same everywhere. The forced "debate the other experts" instruction is gone, and long conversations are trimmed normally (Panel mode never trimmed, which quietly cost more credits). ↻ re-asks your last question on the phone any time, as it already did on the web. New: how to get an independent answer the second model can't be swayed by |
| 5 | 2026-09-17 | Via MCP: a large document that would crowd the expert's context is now held back and reported in the result (documents, mode: held_back, with the cost) instead of being silently left out, and the agent can re-ask with include_documents / exclude_documents. Same behaviour as the app's held-back card, without the pause — an agent cannot answer a card mid-call |
| 4 | 2026-07-23 | Restructured around the two real ways in — conversational "ask for a second opinion" (query_expert, no mode needed) vs. the Panel of Experts toggle's switch-model-and-↻ mechanic — and led with them instead of a defensive "here's what it's not." Corrected "back-and-forth between AI vendors" to "between models" (same-vendor, different-tier comparisons are just as valid). Added the MCP/coding-tool use case: Claude Code (and other MCP clients) can call query_expert mid-work to get a second model's read on a design doc or a stuck code segment. |
| 3 | 2026-07-23 | Corrected core mechanic: Panel of Experts is sequential (switch model → ↻ re-ask), not a simultaneous multi-model fan-out with side-by-side cards. Rewrote What It Does, How to Use It, Tips, Options, and Known Limitations to match; documented the mode toggle's actual location (toolbar circle icon, not inside the Model Selector) and the separate "ask for a second opinion" (query_expert) path available in ambientChat's own chat, not just via Claude Desktop. |
| 2 | 2026-06-07 | Added Web section — Panel mode in the web interface |
| 1 | 2026-03-01 | Initial guide |