The overview says what a council is. This is what it looks like on a Tuesday — the material you hand it, where it sits in your work, how far you turn it up, and the machinery that keeps the result honest.
Anything you would hand a colleague.
Not just code — hand the council a spec, a plan, even a book chapter, plus the criteria that matter to you. Reviewers judge it independently and you get one adjudicated verdict: what to fix first. Flip the ask and the same bench writes the chapter instead.
Seat a local model at the same table as the cloud ones — same prompt, same vote, $0.00, and nothing leaves your machine. Sometimes the free seat is the one that catches it.
Not every ask is “check this.” Point the same bench at the job itself and every seat produces the deliverable — the analysis, the plan, the draft. The judges rank which response best does the work, and the chair merges the strongest into one answer, naming the claims it still rests on.
Monday's council flagged two problems, so you rewrote. Pick up the same thread today — one question, and yesterday's findings flip from red to green.
In the editor, in the pipeline, in front of you.
Release night, a 4,000-line RC diff nobody fully read. Put a council between your pipeline and the tag — exit 0 ships it, exit 1 sends it back.
Configure an MCP server once in Claude Code and every seat picks it up — Gemini, GPT, even your local model. Your tools travel with the council, zero extra setup.
Model catalogs rot quietly — an alias that worked last week can 404 today. One live probe proves every seat on your bench actually serves, so you find out in the check, not mid-council.
Open the Workspace and watch your council think — seats reporting in, phases advancing, costs ticking per model. Finished runs replay end to end.
Every dial is off until you name it.
Models from the same family share the same blind spots — so build a bench that mixes families. Seat as many as you like — two and up — from your own API keys, one OpenRouter key, or local runtimes, each pick earning its chair on specialty, price, or privacy.
The ritual never changes — critic seat, expert lenses, debate rounds, and a Claude seat are all off until you name them. Dial up what you need and see the exact run shape before anything launches.
Every council prints its estimated cost before a single call goes out — and a hard budget cap bounces the run you didn't mean to launch. The receipt matches the estimate, not your worst fear.
The parts that keep a council from being a poll.
Your council ranks every review blind — bylines stripped before the vote, so no model can favor its own. Some have even disputed their own work.
Your plan came from Claude — don't let Claude grade its own homework. A non-Claude chair runs the vote and signs the verdict; Claude only keeps the room in order.
Two reviewers pass your auth refactor; the third catches session fixation. Amicus treats that dissent as the finding — adjudicated and confirmed, not averaged away.
A seat dies mid-run? Every surface tells you — the CLI exits 2, verdict.json flips "degraded": true, and the report wears a red banner. Amicus never degrades silently.