CMS Brew lets your clients edit their own site by chat, safely: the AI resolves every request against a manifest of editable fields built from the repo, so it changes content and never writes code. This package is the developer-facing on-ramp to that.

npx cmsbrew login
npx cmsbrew connect

Commands

Command	What it does
`cmsbrew login`	Store an API key from the dashboard, and verify it works
`cmsbrew logout`	Remove the stored key from this machine
`cmsbrew connect [repo]`	Connect a repository. Defaults to this checkout's `origin`
`cmsbrew sites`	List connected sites and their state
`cmsbrew status [site]`	Show one site in detail
`cmsbrew setup [client]`	Write the MCP config for Claude Code, Cursor, Windsurf, or Codex
`cmsbrew mcp`	Run the MCP server over stdio. Your agent runs this, not you

Global options: --json, --host <url>, --token <key>.

connect

Run it inside a checkout and it reads the origin remote:

cmsbrew connect

Or name the repository, which is what CI and monorepos need:

cmsbrew connect acme/marketing-site
cmsbrew connect acme/monorepo --source-root apps/web --site acme-web

Connecting returns as soon as the site is registered, then CMS Brew maps the repo in the background. The command waits for that and prints the framework and field count when it finishes. Pass --no-wait to return immediately.

MCP

cmsbrew setup detects the MCP clients installed on your machine, shows the config it will add, and asks before writing. It backs up any file it changes.

cmsbrew setup              # every detected client
cmsbrew setup claude-code  # just one

To configure a client by hand, run cmsbrew mcp as the server command with your key in the environment:

{
  "mcpServers": {
    "cmsbrew": {
      "command": "npx",
      "args": ["-y", "cmsbrew", "mcp"],
      "env": { "CMSBREW_TOKEN": "cmsb_..." }
    }
  }
}

Tools

Tool	What the agent can do
`list_sites`	Every connected site, with status, framework, and unpublished-edit state
`connect_site`	Connect a GitHub repository
`get_manifest`	Read a site's editable fields
`edit_content`	Make a content edit in plain language. Lands as a draft
`list_escalations`	Requests routed to the team because they need a developer

There is no publish tool at any permission level. Publishing commits straight to a client's live production site, so it stays a decision a person makes in the dashboard.

Permissions

Every key carries a preset, chosen when you mint it. They nest, narrowest first:

Preset	Can
Setup only	Connect repositories, read sites and manifests
Safe agent access (default)	Setup, plus content edits that land as a draft
Full access	Everything, plus publish and revert

The preset is enforced on the server, so a key cannot exceed it whichever tool calls it. It also cannot exceed the person who minted it: revoke their access to the workspace and their keys stop working.

Keys cannot mint keys. That is deliberate, and it is what keeps a narrow preset meaningful.