Skip to main content
A private tool set is a sub-MCP server mounted inside a running server. It exposes a chosen subset of tools under its own slug and its own MCP endpoint, so one server can present several scoped surfaces instead of one flat catalog.

​
Sub-MCP apps

A sub-MCP app is registered on the live sub-app router under a slug, listing the tools it exposes and the transport it serves. Each registration is a self-contained MCP surface: a client that connects to a sub-MCP sees only that set of tools, not the whole server. The registry keeps a JSON-safe view of every registered sub-MCP — its slug, its tools, and its transport.

​
Registering and tearing down

Sub-MCP apps are managed live, at run time, through three authed doors and their CLI commands:
Registering a slug builds (or reloads) its sub-MCP app; unregistering it tears down the sub-app’s ASGI lifespan cleanly. An unknown slug fails loudly with a 404 rather than silently succeeding.

​
Why scope a tool set

A private tool set narrows what a given client can see and call. Pairing a sub-MCP with access control lets you route different callers to different scoped surfaces of the same server, keeping each client’s tool list minimal and intentional.
The sub-MCP namespace is app.sub_app. It is separate from the raw FastMCP escape hatch on the MCP-citizen surface, which is named app.fastmcp precisely so it does not read as the sub-MCP.
See the private-tool-set guide for exposing one and the HTTP API reference for the /api/sub-mcp routes.