consensus

Throw a hard problem at a panel of frontier models. They answer independently, attack each other's answers, concede or rebut, revise, and repeat until they agree — and you get one answer the whole panel signed off on, with the disagreements that survived laid out honestly.

It runs on the AI subscriptions you already pay for (Claude Pro/Max, ChatGPT Plus/Pro, SuperGrok) as well as API keys or OpenRouter, by driving each vendor's own CLI headless in a clean room. Subscription seats spend the same rate limits as your interactive use of those CLIs, and vendors' terms govern that use — Anthropic's terms restrict third-party products from relying on claude.ai logins. Read this before you rely on it.

# illustrative transcript: a 3-round debate that revises once and converges
$ consensus "Optimistic locking or a distributed lock for inventory holds?"

panel (balanced): claude:claude-opus-5#high, codex:gpt-5.6-sol#high, grok:grok-4.5#high, gemini:gemini-3.8-flash#high
judge: claude:claude-opus-5  rounds: 3  cost: subscription quota; roughly 28 model calls at most
debate log: .consensus/runs/20260916T094012Z-a1b2c3/debate.md  (tail -f to watch)
▶ propose
  ✓ A claude:claude-opus-5   41.2s
  ✓ B codex:gpt-5.6-sol      38.7s
  ...
▶ critique (round 1)
  4 disputes open after round 1
▶ revise (round 1)
▶ critique (round 2)
  panel converged in round 2
▶ synthesize

You get back: the answer, a confidence level, what the panel agreed on, what it could not settle, and what changed during review. Every run is saved and replayable.

Quickstart (90 seconds)

# 1. install: one download (no Node, npm or sudo needed), then the setup wizard
curl -fsSL https://raw.githubusercontent.com/seanheiney/consensus/main/install.sh | sh

# 2. setup connects your accounts, builds profiles, teaches your IDEs, and runs
#    a first debate so you see a real result. Then ask it something:
consensus "Should we move this queue from Postgres SKIP LOCKED to Redis/BullMQ?" -c src/queue.ts

# 3. read it again later
consensus runs          # list past debates
consensus log           # replay the latest in full
consensus log --html    # one self-contained page you can send to someone

Not connected to anything yet? consensus doctor says what is missing and how to fix it. You need at least two seats for a panel; one OpenRouter key (OPENROUTER_API_KEY) is enough to reach every vendor at once.

Want a panel shaped for your problem? Describe it and a model drafts the seats and personas:

consensus profile design "4 panelists: security, distributed systems, a PM, a skeptic; frontier models; 2 rounds"
consensus "…" --profile <the name it chose>

Related MCP server: imladris

Install

The one-liners need no Node, no npm and no sudo: they download a self-contained build (the official Node 22 runtime plus the bundled CLI, about 38 MB), verify it against the release's SHA256SUMS, install it under ~/.consensus with a ~/.local/bin/consensus link, add one marked line to your shell rc files, and start consensus setup in the same terminal. Re-running upgrades in place (and runs consensus doctor instead of the wizard). Your own Node, if any, is never used or changed. No telemetry. Later: consensus update, consensus uninstall --all.

You need at least two model connections for a panel. Installer options: --help, --yes, --no-setup, --version <x.y.z>, --no-modify-path, --dir, and CONSENSUS_* environment equivalents.

Full details — what the installer writes where, verifying releases, proxies, air-gapped installs, upgrading, uninstalling, and a troubleshooting table: docs/install.md.

What consensus setup does

1. Finds your accounts. Vendor CLIs you already use (Claude Code, Codex, Gemini CLI, Grok) and whether they are logged in, plus any API keys. For anything missing it offers to log you in through the vendor's own CLI, install that CLI, or store an API key in ~/.config/consensus/credentials.json (mode 600). Your subscription auth stays inside the vendor's CLI; consensus never sees a token.

2. Builds model profiles from what is connected, and asks which is the default.

3. Teaches your tools. Registers the MCP server and drops a skill pack into Claude Code, Codex, Gemini CLI, Grok, Cursor, Windsurf, Claude Desktop, and the cross-tool ~/.agents/skills directory. Your agent then knows the panel exists and when to reach for it.

4. Runs a first debate so you see a real result and where the log lives.

consensus setup --yes does all of it with no questions. consensus doctor --probe shows what is connected and makes one tiny call through each to prove it. consensus uninstall [--all] [--purge] reverses every change.

How your subscriptions are used

A subscription-backed seat is that vendor's CLI run headless in an empty temp directory as a clean room: Claude Code with no tools, no settings sources, --safe-mode and no MCP servers; Codex with --ignore-user-config, --ignore-rules, a read-only sandbox and its shell/browser/computer-use/app tools disabled; Gemini in plan mode. Panelists see only the prompt you give them — not your files, CLAUDE.md, AGENTS.md, or your other MCP servers.

Read this before relying on subscription seats. Every seat call spends the same rate limits as your interactive use of that CLI (Claude Code's rolling 5-hour and weekly caps, Codex's ChatGPT limits), and a three-round frontier debate can consume a meaningful slice of a daily cap. Vendors' terms govern this use: Anthropic's published terms restrict third-party products from offering or relying on claude.ai logins without approval. consensus is a local tool you run on your own machine under your own login, not a hosted service, but checking your plan's terms is your responsibility — and you should use API keys for anything shared, automated, or run on someone else's behalf. Keys saved by setup are handed only to the matching API client and are never exported into a vendor CLI's environment, so a saved key cannot silently move your subscription CLI onto per-token billing.

Use it from Claude Code, Codex, Cursor…

After setup, your agent has a consensus MCP tool plus a skill telling it when to use it. Just ask: "Get a panel consensus on whether we should migrate this queue to Kafka; include the producer code."

claude mcp add -s user consensus -- consensus mcp      # manual registration, if you prefer
codex mcp add consensus -- consensus mcp
gemini mcp add -s user consensus consensus mcp
grok mcp add consensus consensus -- mcp

Cursor, Windsurf, and Claude Desktop take {"command": "consensus", "args": ["mcp"]} under mcpServers. The server exposes consensus (run a panel) and consensus_profiles (list profiles and connections). See docs/usage.md.

How the protocol works

1. Propose. Every panelist answers independently and in parallel. No anchoring.

2. Critique. Each panelist sees all answers, anonymized as Answer A, B, C (labels shuffled once per run; a panelist knows only which is its own). It must attack every other answer with specific, falsifiable disputes, review its own, and return a verdict per answer: agree (substantively equivalent, no major error) or disagree.

3. Converged? Only when every panelist marks every other answer agree with no major dispute. One dissenter keeps the debate going.

4. Revise. Each panelist gets every dispute raised against it and must concede or rebut each one with a reason, then rewrite its answer.

5. Repeat critique → revise up to --rounds times.

6. Synthesize. The judge writes the unified answer with a fixed structure: the answer, confidence, where the panel agreed, unresolved disagreements with each side's strongest case, and what changed during review.

Design choices: panelists are anonymous so no one defers to a brand; concede-or-rebut is mandatory so critiques cannot be ignored; disagreement is reported rather than hidden; a panelist that errors out is dropped and the run continues while two remain (and the run exits 2 so a script notices); every model gets identical prompts.

With --rounds 1 the panel critiques but never revises — see why.

Cost and time

A 3-model, 3-round run is roughly 20 model calls, each carrying the full set of answers. On subscriptions that is quota; on API keys, expect a few dollars at frontier tier. Every run prints the seats and effort it will use before starting and the estimated list-price cost after; --max-cost 2 aborts mid-run if the estimate crosses a ceiling. A bare single word is refused as a prompt so a typo cannot start a paid run. Use the budget or fast profile for cheap passes.

No benchmark claims are made here: consensus bench is the tool for producing them on your own problems, and the ablation against single-model baselines has not been run yet.

Documentation

Library

import { runConsensus, resolveRun, loadConfig, renderReport } from "consensus-panel";

const cfg = await loadConfig();
const { panel, judge, rounds, effort } = await resolveRun({ cfg, profile: "frontier" });
const run = await runConsensus("Design a rate limiter for a multi-tenant API", { panel, judge, rounds, effort });
console.log(renderReport(run));

Bring your own model by implementing Panelist (id, provider, model, complete(request)). See docs/usage.md.

Development

pnpm install
pnpm test          # protocol, profiles, packs, installer tests; no keys needed
pnpm dev "..."     # run from source
pnpm build && npm link   # `consensus` on PATH from this checkout

See CONTRIBUTING.md. MIT licensed.