dsh-cordis-mcp

Dynamic Cordis plugins in DSH have always been managed through in-session commands like cordis_define, cordis_run, and cordis_stop, leaving no direct hook for external agents. dsh-cordis-mcp bridges the entire plugin lifecycle into MCP (Streamable HTTP) tools so that Claude Code can define Packages, activate them, stop runs, permanently delete plugins, and query internal state—all from a single conversational thread.

On the transport and security side, the MCP endpoint rides on DSHs existing web port with no new listener and no standalone service. Authentication is mandatory Bearer Token; missing or invalid credentials yield 401 or 503. The endpoint accepts only loopback connections (127.*/localhost/::1), rejects cross-site browser requests, and never emits CORS headers. Packages that contain a Client half still go through DSHs native approval flow: dsh_plugin_run returns awaiting-approval and the operator must approve via the Web GUI Run card, preserving the same safety semantics as manual operation.

If you already build dynamic plugins with the DSH Cordis toolkit and use Claude Code as your daily coding and orchestration entry point, this bridge eliminates the repetitive window-switching and hand-typed cordis_* commands. The agent can create, debug, deploy, retire, and inspect plugins end-to-end. Session resolution, multi-live-session disambiguation, and token rotation are handled internally; you simply grab the token from the settings page and wire it into your Claude Code configuration.