← All posts

Let AI agents draft content, and people publish

Connecting an AI agent to a CMS through MCP without giving it the publish button: scoped tokens, approvals it can't give itself, budgets and a kill switch.

Status. Colophon is pre-release (0.1.0-alpha.1) and has no users yet. I build it in public, on my own. See what’s built and the build log.

Content teams want help from AI agents: first drafts, translations, tag clean-ups, link checks. What they don’t want is an agent that can put something on the live site because a prompt told it to. Hiding the publish tool from the agent isn’t enough, because the agent’s credential still reaches the same API. The rule has to live on the server.

What a safe set-up needs

  • Its own credential. One scoped API token per agent, with only the content types and fields it needs, so it can be revoked without stopping anyone else.
  • Drafts only. The agent can write drafts and ask for review, and only people can approve and publish.
  • An approval it can’t give itself. If an agent can submit work and another token can approve it, two agents can still publish together.
  • Stale-write protection. An agent’s edit should land only on the version it read, so it never overwrites a person’s newer work.
  • Limits. Budgets and request quotas, so a looping agent stops instead of running up a bill or a queue.

The full checklist, for any CMS, is in give AI agents CMS access, not the publish button.

How Colophon does it

Colophon’s MCP server, over stdio or Streamable HTTP, offers eighteen tools for reading content, searching it, writing drafts, commenting, filing tasks and requesting review. None of them approves or publishes. That alone would be a convention, so the rule is enforced underneath: the API refuses any token’s approval of what a token submitted, whatever the token’s permissions. Every write starts from what the agent read (update_draft takes the revision it read, request_review the hash it read), so a person’s newer edit is never overwritten. Budgets are reserved in the database per token, with request quotas on top, and revoking the token stops the agent at once.

The rule was tested without a model: an adversarial evaluation suite tried the ways an agent might get content published, and a stored run on 4 November 2026 recorded zero unauthorised actions.

Check your own CMS

Give an agent’s token every permission your CMS allows, then ask it to approve its own submission through the API directly, not through the agent’s tools. If it can, the protection is in the tool list, not in the server.

What to know before choosing Colophon

  • It’s pre-release, and the evaluation is the project’s own; no outside party has tested it.
  • The standalone MCP binary isn’t in the release archives or on npm yet; the server’s own MCP endpoint is included.
  • Agents propose; people still do the reviewing. Colophon makes that enforceable, not faster.

FAQ

Which AI agents work with Colophon?

Any client that speaks the Model Context Protocol, over stdio or Streamable HTTP. Colophon doesn't run a model itself; the agent and its model are yours.

Can an agent publish if I give its token more permissions?

No. There is no publish or approve tool, and the API refuses any token's approval of what a token submitted, so even a token with wide permissions can't approve its own or another token's work.

How do I stop an agent at once?

Revoke its API token. Each agent should have its own, so revoking one stops that agent without touching the others; its next request is refused.