dsh-cordis-mcp

Dynamic Cordis plugins in DSH are managed through in-session commands such as cordis_define, cordis_run, and cordis_stop, but there is no native way for an external agent to drive that lifecycle directly. dsh-cordis-mcp exposes the full plugin lifecycle as ten MCP (Streamable HTTP) tools, letting Claude Code define immutable Packages, activate or update them, stop running instances, permanently delete plugins, inspect providers, and query the tool catalog—all within a single conversation.

The endpoint rides on DSHs existing web port (no new listener, no standalone service). Bearer-token authentication is mandatory: requests without a valid token receive 401 or 503. Access is restricted to loopback addresses only, cross-site browser requests are rejected, and the server never emits CORS headers, making remote abuse infeasible. Packages that include a Client half still follow DSHs built-in approval workflow: the run tool returns awaiting-approval, and the operator confirms via the Run card in the DSH Web GUI, preserving the same safety guarantees as in-session usage.

If you build dynamic plugins with the DSH Cordis toolkit and rely on Claude Code as your daily orchestration surface, this bridge removes the repetitive tab-switching and manual cordis_* typing. Your agent can create, debug, deploy, retire, and inspect plugins end-to-end. Session resolution, multi-session disambiguation, and token rotation are handled internally; you simply copy the token from the settings page and point Claude Code at the endpoint.