Ask several AI models at once. Get one list of everything they raised.

One question in, one list out

1. Ask
2. →
3. Each model answers on its own
4. →
5. One merged list

1. 01

   Ask

   Type a question in your terminal, or use /ai-consultants:consult inside Claude Code. Put file paths after the question. The tool reads them on your machine and saves an exact copy of what it sent as context.md.

2. 02

   Each model answers on its own

   The question goes to every model on your panel at the same time. Each answers once without seeing the others, so there are no debate rounds pulling everyone toward the majority view. Each also gets a different brief (the Architect, the Pragmatist, the Devil's Advocate and so on) so they don't all come at it the same way. Run it from Claude Code and Claude sits this one out: you already have its view.

3. 03

   One merged list

   One more model call merges the answers into a single list of every distinct point, each tagged with who raised it, plus a recommended approach and a risk table with severities and mitigations. A check on your machine then reports whether every finding the models returned is cited somewhere in the merged list (MET, DEGRADED or FAILED) and how many models answered. If a model failed, the report names it and gives the real reason, such as “CLI not found” or “401 Unauthorized”. It checks the bookkeeping, not the content: it can't tell you whether the merged wording is faithful or whether the points are true.

Advice only. The models don't run your code or tests, and nothing in your project changes.

Use it when you want everything listed, not one verdict

Different models often raise different points. That helps most when the job is to list things: risks, edge cases, attack paths, alternatives. It helps little when there's one right answer to find.

Uses the tools and accounts you already have

Pick any two or more. Most are reached through that company's own command-line tool, signed in with your account; GLM and DeepSeek need an API key and are off until you add one; Meta is off until you install its Muse Code CLI or add a key. ai-consultants doesn't sell model access: each company bills your usage to you. Current model versions are listed in the README.

Independent open-source project. Not affiliated with or endorsed by any of these companies.

Cost, time, and where your code goes

Cost

One call per model plus one to merge, and a failed call may be retried up to twice. Calls through a signed-in CLI count against your plan with that company; API keys are billed per use. Every run prints an estimated cost when it finishes, using the rates in docs/COST_RATES.md. Estimates aren't invoices, and models billed by plan or credits are left out (the run says so). A spending cap based on the tool's own estimates is off by default: set ENABLE_BUDGET_LIMIT=true, MAX_SESSION_COST=0.50 and BUDGET_ACTION=stop. It can't see usage billed by plan or credits. Asking the same question again within 24 hours reuses saved answers.

Time

The models answer in parallel, so you wait for the slowest. Expect a few minutes, longer with heavy reasoning models. There's no time limit by default, and every 30 seconds it prints who it's still waiting for. To stop waiting, touch <results folder>/.proceed merges the answers already in; .abort ends the run. A call you stop may still be billed.

Privacy

Your question and every file you name go to each provider on your panel, under that provider's terms. ai-consultants has no server of its own. context.md shows exactly what went out. If your code can't leave your company, don't point it at that code, or limit the panel to providers you're cleared to use.

Set up once, then ask

Setup takes about a minute if two model CLIs are already installed and signed in (longer if you need to set them up). Each question then takes a few minutes.

⚠ Watch for typos: any first word after npx ai-consultants that isn't one of its commands (configure, doctor, init, install, update-clis, version, help) is read as a question and starts a real, billed run.

Questions developers ask first

Why not just ask my own agent?

For a question with one right answer, do. Your agent is one model with one set of habits. Models from other companies, each asked once and separately, often raise points it didn't, and the list shows you which. That's most useful when you're trying to list everything that could go wrong. You still decide what's right.

Is it more accurate than one strong model?

We haven't shown that, and we don't claim it. Two small internal tests found a gain when listing risks and edge cases, and none when hunting planted bugs. The panel shows what several models say and where they agree, and agreement isn't correctness.

Does it vote on one answer?

No. By default it keeps every distinct point and notes who raised it, so a point only one model made isn't averaged away. --strategy compare_only puts the answers side by side with no pick. Other merge styles, including one blended recommendation, are in the docs.

Which models answer, and can I choose?

By default, every model you've set up; you need at least two, or the run won't start. Pick them with npx ai-consultants configure --interactive, or use a mode: --mode fast-check (Gemini and Codex, in cheaper versions), --mode coverage-review (Gemini, Codex and Mistral) or --mode max-coverage (the widest panel at the highest model tier; slowest and most expensive).

Can I use GPT-6.1 Terra?

Yes. Version 5.7.0 adds it as an optional Codex model: set CODEX_MODEL=gpt-6.1-terra in your configuration. Terra remains opt-in. Version 5.8.0 refreshes the Claude and OpenAI tier choices following the AA v5 review; explicit effort settings still win. Leave the effort unset to keep the transport default, or choose low through max; API mode also accepts none. See the v5.8.0 release notes and pricing and configuration details.

What does a run cost?

One call per model plus one to merge, billed to your own accounts, and a failed call may be retried up to twice. Every run prints an estimated cost when it finishes. Asking the same question again within 24 hours reuses saved answers.

Where do the results go?

To a folder under ~/.cache/ai-consultants/consultations/. It holds report.md, one JSON file per model with its full answer, synthesis.json with the merged list, and context.md with exactly what was sent. Inside Claude Code, the command also shows a summary of the same report.

Does it change my code or run anything?

No. It's advice only: the models don't run your code or tests, and nothing in your project changes. You decide what to do with the list.