local-mcp-file-editing

Run local-mcp mcp to start the server, then run local-mcp start in your project directory to create a session. Give the printed session ID to your agent, which can then call read_file, write_file, and list_directory scoped to that working directory.

Use the execute tool with a session_id, command, and args. Commands run without network access via Landlock on Linux or Seatbelt on macOS; commands needing network access go through without_sandbox, which requires explicit approval unless yolo mode is set.

No, local-mcp does not yet support Windows. It runs on Linux, which requires bubblewrap and the codex-linux-sandbox helper binary, and on macOS, which uses the built-in sandbox-exec with no extra dependencies.

Commands fail when they touch paths outside the session working directory or need network access. Check the mode with /permissions status, allow extra directories with /permissions allow, or use without_sandbox with approval for network operations.

Use start_command to launch the process in the background and receive a job_id immediately. Check its status with poll_job and terminate it with stop_job when finished; commands over 30 seconds via execute also return a job_id automatically.