CMS Brew
Start for $0

Your client sites, run by an agent.

Connect Claude Code, Cursor, or any MCP client and run the whole CMS by prompt. Agents stage edits. People publish.

claude - cmsbrew mcpconnected

Connect acme/marketing-site and tell me what my client will be able to edit.

connect_site repo: acme/marketing-site

Scanning repo structure...

Detected: astro

29 editable fields mapped

Site ready: cmsbrew.com/dashboard?site=acme

Remote MCPhttps://cmsbrew.com/api/mcp

Works with Claude Code, Cursor, and any client that accepts a URL and a header. Claude.ai needs OAuth, which this endpoint does not do yet.

Try the live demoStart a 7-day trial for $0

No credit card, cancel anytime.

Connected in three commands.

You run these once. Your client only ever sees the chat editor on the far side.

  1. 01

    Sign in

    Mint a key in the dashboard and store it. The command opens the page for you.

    npx cmsbrew login
  2. 02

    Connect a site

    Run it inside a checkout and it reads the origin remote. Name the repository instead for CI or a monorepo.

    npx cmsbrew connect
  3. 03

    Connect your agent

    Detects the MCP clients on your machine, shows the config it will add, and asks before writing.

    npx cmsbrew setup

Works with Claude Code, Cursor, Windsurf, and Codex.

Then paste your first prompt

You're connected to CMS Brew. Call list_sites to see my sites, then get_manifest on one of them and tell me what a client could safely edit there.

Read-only, so it works with any key preset and changes nothing.

Five tools. That is the whole list.

Enough for an agent to onboard a site and keep its content current. Nothing that reaches the deploy.

  • list_sites

    Every connected site, with status, framework, and unpublished edits.

  • connect_site

    Connect a GitHub repository and map its editable fields.

  • get_manifest

    Read the fields a site exposes, and the guard tier on each one.

  • edit_content

    Make a content edit in plain language. It lands as a draft.

  • list_escalations

    Requests routed to the team because they need a developer.

An agent that edits the site and still cannot publish it.

The CLI and the MCP server reach the same API the dashboard uses. Every key carries a preset, chosen when you mint it and enforced on the server. The default is the narrow one.

  • Setup only

    Can: Connect repositories, read sites and manifests.

    Cannot: Edit or publish.

  • Safe agent access

    Default

    Can: Everything above, plus content edits that land as a draft.

    Cannot: Publish to the live site.

  • Full access

    Can: Everything above, plus publish and revert.

    Cannot: Act outside the workspace it was minted for.

  • No publish tool. At any tier.

    Publishing commits to a live production site, so it stays a decision a person makes in the dashboard. The agent stages the work and stops.

  • Keys cannot mint keys.

    Minting is closed to every key and open only to a signed-in operator. A narrow preset is one call away from meaningless the moment that is not true.

  • A key never outranks its owner.

    It acts as the person who created it. Remove them from the workspace and their keys stop working on the next call.

Stop being the person who updates the phone number.

Give your client the keys to the content. Keep the keys to the code.

Start for $0

Prefer to talk first? Book a 30-minute call