Skip to content
中文

MCP servers

MCP (Model Context Protocol) connects external tools and data sources. Once a server is configured, its tools appear in the agent resource list and can be selected.

Type Use when Key fields
SSE Remote service that pushes events Endpoint URL, headers
HTTP Remote request/response service Endpoint URL, headers
WebSocket Long-lived bidirectional connection Endpoint URL, headers
Stdio A local executable Command, arguments, environment

Remote types need an endpoint URL and optional headers. Stdio needs a command and arguments.

  1. Open MCP → New server.
  2. Name it by capability, e.g. filesystem or internal-search.
  3. Description: explain what it provides so it is recognisable in the agent editor.
  4. Type: choose from the table above.
  5. Connection: URL for remote; command and one-argument-per-line for Stdio.
  6. Timeout (seconds): raise it for slow tools.
  7. Save. The platform discovers tools; the list appears in the server details.

Open the Agents editor and tick the server under MCP servers. Unticked servers are never called by that agent.

Connection fails — test/refresh in the server details; verify URL, headers and timeout. For Stdio, confirm the command exists in the runtime.

No tools listed — the server may expose resources but no tools, or the handshake failed. Check logs.

Tool calls time out — raise the timeout, or use an asynchronous task-style tool.

Several agents, one server — configure once and tick it in each agent.