External MCP servers plug straight into the agent's tool layer.

Every server tool appears as mcp_<server>_<tool> — e.g. mcp_github_create_issue — and is permission-gated like any built-in tool.

1. A server is registered (URL or command + capabilities)
2. Its tools join the model's tool list automatically
3. The model calls them like native tools; results flow back into the chat

Issue trackers, databases, browsers, home automation — anything with an MCP server becomes a tool the agent can call, without shipping new app code.

---