amicus curiae (n.) — one who is not a party to the case, but is permitted to advise it.
A multi-model AI council and ad-hoc sidecar, with Claude as orchestrator.
Three convenings. Different seats, different chair, different work.
The ritual doesn't change — and a model on your own machine is a full member.
Asking Claude twice gets you the same answer twice. Asking four models from four different families gets you the disagreement — which is the part worth reading.
Models trained alike miss alike. A bench that mixes Gemini, GPT, DeepSeek, and a local Llama doesn't fail in unison.
Two reviewers pass your auth refactor; the third catches session fixation. Amicus adjudicates that split and surfaces it — instead of averaging it into nothing.
Claude wrote the plan, so a non-Claude model chairs the vote and writes the verdict. Claude keeps the room in order — it never sits as judge.
Anything you'd hand a colleague — check my plan, write the migration, choose between these two designs. They work it independently, rank each other's answers without knowing whose is whose, and one result comes back with the disagreements still in it.
In Claude Code or Cowork. Amicus picks up the material and your whole session — no exporting, no pasting.
Same doc, same criteria, same context. No reviewer sees another's work while writing their own.
Reviews swap hands with the bylines stripped. Models rank the argument, not the reputation.
One ranked document — what they agreed on, what split them, what only one of them saw — folded straight back into Claude's context.
You hand them your work. They tell you what's wrong with it — every defect tiered by how many seats confirmed it.
sessions.user_id1/3You hand them the problem. They do the work — and the claims the answer rests on get tiered exactly the same way.
Either way what comes back is not three opinions. It is one account of what they agreed on, what split them, and what only one of them saw — and the tiers report peer concurrence, never verification. Three models agreeing are still three models agreeing.
Sometimes you just want a second pair of eyes. Name a model and Claude hands it the conversation — your full context already loaded, so there's nothing to re-explain. Work alongside it, then say fold and a structured summary flows back into Claude.
amicus start --model gemini
A single OpenRouter key puts 200+ models on tap — Google, OpenAI, Anthropic, xAI, Meta, DeepSeek, Mistral, Qwen — with routing, fallback, and billing already handled. They even publish free variants, so your first council can cost nothing at all.
That's what makes a real council practical. Swapping DeepSeek for Qwen is a word change, not a procurement exercise. Want seven seats from seven different families? You already have them. OpenRouter does that heavy lifting — Amicus just seats the table.
Already have your own keys? Direct Google, OpenAI, Anthropic, and DeepSeek keys work exactly the same — and local models via Ollama or LM Studio need no key at all.
One installs the CLI, both Claude skills, and the MCP server. The other adds your keys. Amicus can't reach a model until amicus setup has run.
node --version to check.Amicus does more than fits on one page. Here's the rest in a line each — and fifteen of them drawn out if you'd rather see than read.
degraded, and the report wears a banner.Install in 30 seconds. Convene your first council in one sentence.