mcp_servers.yaml and loaded at startup; their tools are registered
under the server name as a namespace and enter the same progressive discovery
surface as native tools, so the model finds them with search_tools and
activates them with load_tools. A server that fails to connect is non-fatal:
the rest of the session continues.
It will cover the four transports (stdio, sse, streamable_http, and
websocket), OAuth acquisition and refresh including the device-code flow used
when running headless, and the live MCP manager that holds one persistent session
per server so repeat tool calls reuse an open connection instead of reconnecting.
Finally it will cover the control panel: Ctrl+O or /mcp in the TUI opens a
table of servers with state, transport, auth, and tool count, mapping connect,
disconnect, and reconnect to the manager, with a tool peek and an OAuth prompt
modal. In the REPL, /mcp [server] prints the same information as a table.
This page is in progress. Until it lands, the shipped
mcp_servers.yaml files
in each agent folder are the best worked examples.mcp_servers.yaml resolves.