Skip to main content

​
buttons mcp

Run an MCP server over stdio (expose buttons to agents)

​
Synopsis

Start a Model Context Protocol server on stdio so an agent (e.g. Claude Code) can discover and press buttons as tools. Uses a thin meta-tool surface — buttons_list, buttons_press, buttons_inspect (and buttons_create with —allow-create) — instead of one tool per button, so large button sets don’t degrade the MCP client. Security:
  • Only buttons with “mcp_enabled”: true are listed, pressable, or inspectable.
  • Args are validated against the button’s spec and passed as BUTTONS_ARG_<NAME> env vars — never substituted into shell text.
  • Per button: max 10 calls/min, 1 concurrent press, hard 120s timeout cap.
  • buttons_create is OFF unless —allow-create is passed.
stdout carries only protocol messages; logs go to stderr. Register with an agent, e.g. Claude Code: claude mcp add buttons — buttons mcp 
buttons mcp [flags]

​
Options

      --allow-create   expose the buttons_create tool (off by default)
  -h, --help           help for mcp

​
Options inherited from parent commands

      --json       output in JSON format
      --no-input   disable all interactive prompts
      --summary    show a read-only plan/snapshot instead of mutating

​
SEE ALSO

  • buttons - Deterministic workflow engine for agents