mcp_install¶
Guided install of an MCP server from the registry into the project configuration.
Entry¶
discover
Final output¶
mcp_install_result — status ("installed" / "error" / "skipped"), server_id, server_name, scope, installed_path.
How it composes¶
A single discover phase. The registry_fetch.py preprocessor (unsafe mode, 20 s timeout) queries the MCP registry and populates data.registry before the LLM runs. The LLM reads data.registry.source (direct / search / not_found / error), optionally calls ask_user to resolve ambiguity, then emits one mcp_install Control IR op. The OS layer owns the runtime permission check, secret prompting, config writing, and event emission.
Caveats¶
Requires mcp_install: true in reyn.yaml permissions. Scope defaults to local (writes to reyn.local.yaml). Anthropic reference servers (GitHub, filesystem, memory) are not listed in the registry — provide an explicit server_id for those.
Usage¶
reyn run mcp_install "filesystem MCP server を入れて"
reyn run mcp_install '{"server_id": "modelcontextprotocol/server-github"}'