MCP Command¶

The mcp command exposes your CLI tool's functionality to AI agents using the Model Context Protocol (MCP).

Usage¶

mytool mcp [subcommand] [flags]

Description¶

Starts an MCP server that allows AI assistants (like Claude or Gemini) to discover and execute your CLI's commands as tools. This enables agentic workflows where an AI can autonomously use your tool to perform tasks.

Subcommands¶

`mcp start`¶

Starts the MCP server over standard I/O.

mytool mcp start [--debug]

Flag	Description
`--debug`	Enable debug logging for MCP communication

`mcp stream`¶

Starts the MCP server over a networked streamable-HTTP endpoint, serving the same tools as mcp start but over HTTP instead of standard I/O.

mytool mcp stream [--host <host>] [--port <port>]

Flag	Description
`--host`	Host/interface to bind the HTTP listener to
`--port`	Port to serve the streamable-HTTP endpoint on

Use mcp start (stdio) for editor and desktop integrations that launch your tool as a subprocess (e.g. Claude Desktop); use mcp stream for remote or networked clients that connect to a running HTTP endpoint rather than spawning the process locally.

`mcp tools`¶

Exports the tool definitions to a JSON file for inspection.

mytool mcp tools

This generates an mcp-tools.json file in your current directory, showing the JSON schema for each exposed command. This is useful for:

Debugging which commands are exposed to AI agents
Understanding the expected input/output format
Validating tool definitions before deployment

Gating sensitive commands¶

Every command in the tree is on the MCP tool surface by default. A command that should stay runnable on the CLI but not be callable by an agent is excluded at build time:

setup.ExcludeFromMCP(mycmd.NewCmdDeploy(props))

Exclusion is inherited: descendants of an excluded command are excluded too, unless one of them opts back in.

setup.IncludeInMCP(mycmd.NewCmdDeployStatus(props))

IsExposedToMCP walks a command and its ancestors and takes the nearest explicit decision — exposed wins for that command, excluded hides the subtree, and a tree with no decision anywhere is exposed.

The generator carries the same tri-state through .gtb/manifest.yaml:

gtb generate command -n post --mcp-enabled=false

Omitting the flag leaves the decision unset (inherit); --mcp-enabled=true records an explicit exposure, which is how you re-expose one subcommand of an excluded parent.

Exposure is build-time only¶

There is no configuration key or flag that changes the MCP surface at runtime. The decision is baked into the binary as a cobra annotation (gtb.mcp.exposure), so the surface a shipped binary presents is fixed and auditable. An operator cannot widen it, and a compromised config file cannot either.

Confirm what a given build actually exposes with mytool mcp tools, which writes the full tool definitions to mcp-tools.json.