> ## Documentation Index
> Fetch the complete documentation index at: https://docs.qredence.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP integration

> Register Model Context Protocol servers so their tools show up alongside Fleet Prime Agent's built-ins.

Model Context Protocol (MCP) is how you add external tool servers to the agent. Fleet Prime Agent treats every registered MCP tool as a normal `AgentTool` — the model can call it, the runtime renders it as a card, and you don't need any custom glue on the agent side.

## Register a server

MCP servers are exposed through the extension runtime (`packages/coding-agent/src/core/extensions/`). Register an MCP-capable extension the same way you register any other extension:

```bash theme={null}
prime-agent --extension @your-org/mcp-github
```

Or add it to your project or personal extension config so it loads every time. See [Extensions](/fleet-prime-agent/guides/extensions) for the config paths.

## How it looks to the model

Once registered, the MCP server's tools appear in the model's tool list. Each tool implements the standard `AgentTool` interface (`execute`, `prepareArguments`) so the model treats it identically to a built-in tool. Tool cards render the same way as [tool cards](/fleet-prime-agent/guides/tool-cards) for shell or IPython.

## Capability gating

If an MCP extension affects daemon behavior (not just adds tools), it must sit behind a negotiated capability gate. The daemon protocol won't send new commands to clients that didn't advertise the capability, so mixed old-client / new-daemon setups keep working. See [Daemon protocol](/fleet-prime-agent/reference/daemon-protocol).

## Related

* [Tools, skills, and extensions](/fleet-prime-agent/concepts/tools-and-extensions) — the three-tier capability model.
* [Extensions](/fleet-prime-agent/guides/extensions) — the general mechanism MCP builds on.