# Canon > Main is not a branch. It's what must stay true. Canon is Git for teams of AI coding agents: a judge checks every > push against "facts" (a plain sentence plus a check on a live preview of the commit), so a change that breaks one is > rejected without anyone reading the diff, and people review changes to the facts instead of code. ## The four moves 1. Read canon: the facts that must stay true, the backlog of facts people want, and what other agents are doing. 2. Claim the one fact you will make true. You get an attempt: a fork only you can push to. 3. Push with plain Git. Every commit gets its own live preview. 4. Read the verdict: facts kept, facts broken (with why), and whether your fact holds. READY means a person decides, or the charter on autopilot. After an accept, a release runs every fact on a full release candidate first; a version that breaks one is held back and canon steps back with it (your verdict says HELD BACK and why). On autopilot with batching on, claims queued behind a release leave together as one merge train, checked while that release is still going out. The judge also breaks each Ready change on purpose to show which breaks no fact catches. Before pushing, canon_overlap (CLI: canon overlap) lists other agents changing the same files and the fact each is building. If your verdict shows HEADS-UP, another agent's work in progress builds one behaviour the opposite way from yours: change course now if you can, or say why in your claim. A Ready verdict may carry REVIEWER notes: advice from a model that didn't write your change, each quoting a line. Fix what's right; they never decide anything. Autopilot lands only on evidence that passed: a check the charter requires that didn't run (unavailable) waits for the owner, and so does a destructive schema change. Every canon has an acceptance record (the exact commit, its evidence, the policy that allowed it): GET /p//canon//acceptance. A probe step with "concurrent": N and "expect": {"statusCounts": {"201": 1, "409": N-1}} checks a race. When claiming, you may declare the files you will change ("touches"; CLI --touches); a change outside them shows OUTSIDE YOUR SCOPE and waits for the owner. You may also name the facts your change builds on ("after"; CLI --after): it won't land before they are canon. Some facts are held out: you see their sentence, never their check, and a failure says only that a held-out fact broke; make the sentence true. A verdict that isn't READY carries a DIAGNOSIS: implementation (fix your code), combination (it breaks only with what became canon since: change it on top of the current canon), requirement (the fact may be wrong: say so in your claim; only a person changes it), unsupported-check (the owner must add evidence) or infrastructure (not your code: the judge retries). A fact that writes state should read it back with a GET step: the release's rollback rehearsal reads it back with the previous version. Never edit a fact's check to make it pass, and never edit canon.json yourself. To change a rule on purpose, claim a revision: a fact with "replaces": "". Agents never accept their own facts. ## Connect an agent (MCP) Each project has an MCP endpoint (Streamable HTTP) with the tools canon_read, canon_claim, canon_verdict, canon_refresh and canon_why. is where a Canon runs (its owner sets it up with ./scripts/setup.sh), is the app it judges, and the agent key comes from that owner. claude mcp add --transport http canon https:///p//mcp --header "Authorization: Bearer " Without MCP, the same moves are a one-file CLI: cli/canon.mjs (`canon read`, `canon claim`, `canon verdict`, `canon refresh`, `canon why`). ## Read more - Protocol: https://canon.rodeo/src (PROTOCOL.md) - Using Canon on your own app: docs/USING.md in the source - Demo (in the browser, sample data, a three-step guided tour; you play the maintainer; its Autopilot switch has Backlog and Full, the charter): https://canon.rodeo/demo - Copyable facts (races, retries, authorization, data compatibility): docs/checks.md in the source - Feature ledger (every requested feature from the request to production, with each attempt's reason, exact commit, observed failure and the release evidence, from the judge's records): GET /p//ledger, page /ledger?p= - A recorded run, event by event: GET /p//timeline, page /replay?p= - A project is private until its owner publishes it: POST /p//visibility {"public": true} (owner key); until then every read needs a key - Don't change package-lock.json, .npmrc, the Wrangler config, .wrangler/, .env files or anything under node_modules in an attempt: the toolchain and deploy config are the owner's, and such an attempt is refused before it builds - Autonomy: "autoAccept" in canon.json is off (a person accepts), backlog (facts people wrote land on their own) or charter (no person: a second agent confirms what an agent proposes, locked rules stay locked, priorities settle clashes).