MCP server

@cronwatch/mcp is a Model Context Protocol server over stdio. It talks to the JSON API your app mounts with cw.routes(), so it needs a URL and the token, and nothing else.

Claude Code

claude mcp add cronwatch \
  -e CRONWATCH_URL=https://app.example.com/cronwatch \
  -e CRONWATCH_TOKEN=your-token \
  -- npx -y @cronwatch/mcp

Other clients

Any client that launches stdio servers:

{
  "mcpServers": {
    "cronwatch": {
      "command": "npx",
      "args": ["-y", "@cronwatch/mcp"],
      "env": { "CRONWATCH_URL": "https://app.example.com/cronwatch", "CRONWATCH_TOKEN": "your-token" }
    }
  }
}

--url and --token work as flags too. For a local app leave the token out and the routes open in development.

Tools

Tool	Does
`list_jobs`	every job with health, schedule, last run and next due. The place to start.
`get_job`	one job in detail: definition, open conditions, and recent runs with errors, output tails and metrics
`run_check`	look for missed and stuck runs now and send due alerts
`silence_job`	stop alerts for a duration, for example during a fix
`unsilence_job`	resume them
`forget_job`	remove a job that no longer exists in the code
`get_setup_guide`	the code to add CronWatch to a job, so the agent writes it correctly

The tools return prose an agent can act on, not raw JSON. A typical exchange: "why did invoice-run fail last night" becomes get_job, a read of the error and the earlier runs, and a suggested fix in your code.

Security

The token gives the agent everything the dashboard can do. Give it a dedicated app deployment or accept that an agent can silence and forget jobs. Run history includes whatever your jobs logged; if that is sensitive, log less.

The agent skill

A Claude Code skill in the repository teaches the workflow: where to declare jobs, how to wrap them, how to investigate a failure with these tools. See Agent skill.