Ask one model whether to use a queue and it says yes. Ask five, naively, and all five say yes — one opinion wearing five hats. Quorum seats a table of AI specialists on your own machine, makes them argue it out, and hands you a decision record with the dissent that survived.
Runs on 127.0.0.1. Bring your own Claude account. macOS, Node 20+, MIT-licensed core.
The roundtable
Seat two specialists in the free edition, up to five with Pro, and pose a question about your codebase. Openings are written blind and in parallel, so nobody anchors on whoever spoke first. Then everyone has to name the strongest argument against themselves — and rebut it, or concede out loud.
The cockpit
Claude Code sessions, background jobs, Codex, Hermes, ComfyUI renders and memory pressure on one page, with real terminals in a drawer. Projects are discovered from your workspace, not configured — a seven-step tour runs on first launch and there is nothing to set up before it works.
Local-first
Binds to 127.0.0.1 and nowhere else. Debate turns run as headless claude -p calls on your signed-in Claude Code CLI, or an API key already in your environment — Quorum never collects, stores, shows or logs one. No servers, no account, no token resale.
Four phases. Each turn is a separate headless call with tools and MCP stripped — debaters reason, they never touch your files. The turn count and a cost estimate are shown before anything is spent, the running total is live, and cancel kills the child processes.
Phase 1
Written in parallel and blind. No participant sees another's, so positions are independent rather than anchored.
Phase 2
Everyone sees every opening and must name the strongest argument against themselves, then rebut it concretely or concede. Conceding is recorded, not penalised.
Phase 3
A final position and what changed. Confidence is tracked across phases, so a mind actually changing is visible.
Phase 4
The moderator writes the decision, the dissent that survived, and what is still unknown. Finished debates persist locally and export as a markdown decision record.
Each persona is given something different to lose. That is what produces friction instead of agreement. The free edition runs real debates: Vex and Bolt are directly opposed — long-term cost against shipping today — and Nib moderates.
| Name | Role | Argues from | Edition |
|---|---|---|---|
| Nib | Moderator | frames the decision, writes the record, never takes a side | Free |
| Vex | Architect | two-year coupling cost, blast radius, boundaries | Free |
| Bolt | Builder | shortest path to something running and testable today | Free |
| Sable | Adversary | paid to break it — concrete failure modes, not vague worries | Pro |
| Muse | Designer | the person holding it, naming, defaults, silent failures | Pro |
| Ledger | Operator | monitoring, rollback, cost, who gets woken at 3am | Pro |
The open core is free and always will be. Everything below is a one-time purchase — there is no subscription, because there is nothing to host. Debate turns run on your own Claude account, so the only recurring cost is the one you already pay Anthropic.
Tier one
$99
The full six-character crew, and the ability to write your own.
~/.quorum/cast/Tier three
$129
Quorum Pro and The Quorum Method together. The intended purchase.
Tier two
$79
A six-module course on running a roundtable that is worth the tokens — currently in production.
30-day refund, no questions asked — on every paid tier. A refunded licence is not revoked; it verifies offline, and the cost of the odd freeloader is smaller than the cost of building revocation.
macOS, Node 20 or newer, git, and the Claude Code CLI signed in. The cockpit runs without Claude Code; the roundtable doesn't. If you would rather use an API key, start Quorum with ANTHROPIC_API_KEY in the process environment — it is read from there and never written anywhere. Linux and Windows are not supported yet. Whether your Claude subscription permits headless claude -p use is set by Anthropic’s terms for your account; an API key is always a supported path, and model usage is billed by Anthropic, never included in any Quorum price.
Only the debate turns you start, which go to Anthropic on your own account exactly as a Claude Code session would. The server binds to loopback and nothing else. No analytics, no telemetry, no account with us. The bootstrap that proposes a setup reads which directories your sessions run in — transcripts are never sent.
Each turn is a separate headless call with tools and MCP stripped, which makes it a fraction of a tooled Claude turn. Quorum shows the turn count and a cost estimate before you start and keeps a running total while it runs. That cost is billed by Anthropic to you; none of it comes to Trident Studio.
No. Nib, Vex and Bolt run a real two-sided argument — long-term cost against shipping today, with a moderator who never takes a side. Pro adds the three seats that are greyed out at the table and lets you write your own.
It is a signed file the Pro installer places at ~/.quorum/licence.json. Verification is offline — there is no activation server to go down and nothing that phones home. The licence is perpetual for the version you receive; how updates are delivered after the first year is still being decided.
The listings are being finalised. Until the buttons above go live, the open core is on GitHub and the free crew runs full debates today. If you want to be told when Pro ships, leave your email.
Run your first debate today.
Clone it, start it, seat the free crew, and ask the question you have been going back and forth on. It persists the record whether or not you ever buy anything.
Quorum is built and run by Trident Studio in Fort Collins, Colorado. The studio runs on it daily. Any debate we publish is a real recording — a concession is never staged.