Canon in the loop — scaffold, check, publish from MCP

Most MCP demos end when the tool returns JSON.

HUMΛN’s Developer Tools + Canon loop starts when the tool returns judgment about the work itself: is this doc Canon-shaped? Does this scaffold match the operating model? Can we sign and publish without smuggling a fifth status ledger into notes/?

The Friday failure mode

An agent writes a “KB update,” dumps it in chat, and you paste it somewhere. Monday, nobody knows whether it was Canonical, Draft, or fanfic. The agent was helpful. The system was not.

Canon-as-Code exists so operational truth is versioned and reviewable. MCP exists so the agent that built the change can check it before you merge.

The loop (scaffold → simulate → sign → publish)

From an MCP client with the right delegation:

Scaffold — human.canon.scaffold_doc (or CLI equivalents) produces structure that passes frontmatter and Layer D expectations.
Simulate / check — human.canon.check runs the same class of gates engineers run in CI (kb:validate, compliance shapes — not “looks fine to me”).
Sign / provenance — marketplace and publish paths expect signatures and receipts, not vibes.
Publish — into the surfaces that already have gates (docs refresh, public surfaces program), not a parallel markdown graveyard.

You stay in Cursor or Claude Code the whole time. The Companion-as-colleague path (human.ask) can drive the narrative; the Canon tools enforce the bar.

So that agents stop being unpaid interns with a delete key

Governance that only lives in a human’s head does not scale. Governance that only lives in a CI job the agent never sees gets bypassed. Putting Canon check on the same wire as tools/call means the agent that proposes the change can fail closed before you review.

Scroll-stopper: If your coding agent can ship product but cannot check Canon, you automated the easy half of the job.

What this is not

This is not “let the model invent company policy.” HumanOS still enforces policies you define. Canon MCP tools enforce documentation and operating-model shape — the memory the org agrees to maintain — so agents and humans share one ledger of truth.

Go deeper

Product: Developers
Docs: MCP integration, CLI reference
Community: Org mesh + catalog, Resources + prompts, Getting started