Model Council

English | 简体中文

An MCP server that seats other LLMs at your table. Your assistant asks them, reads their answers as tool results, relays those answers back and forth for critique, and gives you one merged conclusion — inside a single normal conversation, with no copy-paste.

Your assistant chairs the council. Any number of members, from any mix of OpenAI-compatible and Anthropic-compatible endpoints — a hosted API, a self-run gateway, a local server, or several of each.

Tools

Members are stateless and cannot see your conversation, so the chair passes everything they need in each call. That is exactly what makes cross-review work: it puts one member's answer inside another's prompt.

Giving the council something to read

A question is usually about something — a spec, a log, a diff, a screenshot. The obvious way to include it is to paste it into prompt, and that is the expensive one, in the place nobody watches: prompt is an argument the chair writes, so a long document costs a full copy of itself in generated tokens on every call, and what reaches the members is whatever the chair managed to reproduce. For forty pages that is not reliably the document. A council reviewing a paraphrase is not reviewing the thing, and nothing in the transcript would say so.

materials names it instead:

ask_all(
  prompt="Where would this break under load?",
  materials=[
    {"label": "The design", "path": "/abs/path/design.md"},
    {"label": "Latency graph", "path": "/abs/path/p99.png"},
    {"label": "My summary", "text": "…something with no file behind it…"},
  ],
  rounds=2,
)

The server reads each one and puts it ahead of the question, in the same place for every member and every round. The position is not presentation:

• Identical bytes. Every member argues with the same copy, which is what makes their disagreement about the document rather than about which version each was given.
• A prefix that can be cached. Everything up to the end of the material is the same in every call of a discussion. The Anthropic format is told so explicitly, with one cache_control breakpoint at that boundary; endpoints that cache by prefix on their own get the shape they need either way. Set "cache": false on a member whose gateway rejects the field.
• The chair pays once. A path costs the chair a line, not a copy of the document, and ask_all carries the material into round 2 and 3 by itself.

Images are the case that matters most. Without them the chair has to describe the screenshot in prose — and then every member reads the same description, so anything the chair misread is misread by the whole council at once, and cross-review cannot recover it. It is the one input where passing it badly quietly removes the independence the council is for. Png, jpeg, gif and webp go to every member whose model can see; a member that cannot should be configured "vision": false, and it then sits out the calls that carry an image rather than answering about text alone as though it had seen the picture. The transcript names who sat out.

How to tell which members those are: a member that cannot see rarely says so. One Anthropic-compatible gateway tested while building this accepted the image blocks, discarded them, and the model then named a colour — not "I was sent no image", which the prompt had explicitly asked for in that case. Text material through the same endpoint arrived intact, so nothing about the call looked wrong. Show a member one unambiguous picture and ask what is in it; a confidently wrong answer is the symptom, and "vision": false is the fix.

Two limits worth knowing. A material that cannot be read stops the call — a council asked about a document it never received will answer anyway, fluently, and read exactly like one that had. And revision_prompt names your files rather than pasting them back, at the position the members were given them, so open them for your own seat before you hand it the rest.

Whether this server will read a path at all depends on how it was started; over HTTP it will not, unless its operator named a directory. list_council says which, in a line above the table.

Rounds

ask_all(prompt, rounds=2) runs that cross-review for you. Round 1 is the usual parallel ask. Round 2 goes back to each member carrying the question plus every answer from round 1 — its own and the others', verbatim — and asks it to revise: take what is right, correct what is not, and say where it still disagrees and why. The transcript comes back round by round, so you can see who moved and who held their ground.

Carrying the previous round back is the whole mechanism. Members remember nothing between calls, so without it a second round is just the same question asked twice. Up to 3 rounds; each one costs another call per member and a longer prompt than the last, so 1 is right for a survey of opinion and 2 for a question where the disagreement is the interesting part.

rounds is chosen before anything has been asked, which is the one thing wrong with it: you commit to a second round without having read the first, and a council that turns out to agree costs exactly what one still arguing would. Ask with rounds=1 when you would rather look first, then buy the next round with revise — it runs exactly one more, as many times as you judge it worth, and has no ceiling of its own. Reading before you buy is usually the cheaper side: another round is a full call per member, while revise costs you only the answers written back into it. The 3 on ask_all is a ceiling on what one call will spend, not on what the discussion can have, and the transcript says so when you reach it.

Members can also carry different weights, for the common case where the council is not a council of equals.

The standing objection

A council mostly agrees, and its agreement is the least informative thing it produces. Everything above works against convergence — the members are anonymous to each other, they are never told the weights, and the closing instruction tells them not to cave — but none of it creates any pressure to diverge. The strongest objection to a plan is not volunteered by members who think the plan is fine.

steelman seats one:

ask_all(prompt, rounds=3, steelman={})

One member writes the strongest case against whatever the table has converged on, each round, and it goes back to everyone as an ordinary anonymous answer. In the next round they have to deal with it.

No member is told to argue a side it does not hold. That is the line, and it is what separates this from a debate mode. What the members say is still what they think; the assignment lives in one extra call they are never told about. Handed a seat marked "arguing against", a model discounts the argument instead of answering it — so the provenance goes to you, in a note at the end of the transcript, and the argument goes to them.

It speaks every round, not once. This is the part that looks like a detail and is not. An objection that cannot reply to its own rebuttal is quoted rather than represented: it cannot correct a misreading of itself, so by the third round the table is arguing with its paraphrase and calling that an answer. tenure buys fewer rounds than the default, and when the seat is retired early the transcript says it was retired by configuration — because an unexplained silence reads exactly like a position abandoned.

The mirror of that is worth as much: the last thing the objection says is the last thing in the transcript, so nobody has been asked to take it on, and the transcript says that too. An objection nobody answered is not a point that stood — it is a point that was never examined. With rounds=2 that is true of everything it said, which is why the note names the setting that fixes it rather than leaving you to notice after paying.

Read what it produces as the strongest objection this council can make to order, never as evidence that anyone holds it. A point of its that survives being answered is worth something; the same point in the round it appeared is worth nothing yet.

By default the first member being asked writes it, and also answers as itself — the two calls are unrelated as far as it knows. Name another with steelman={"model": "glm"}.

Seating an answer you already have

Your assistant is not only the chair — it can answer too, and in Claude Code or Codex it can spawn a subagent to answer as well. Those answers used to sit beside the council's, compared by hand at the end. guests puts them in it: