dsh-mcp-manager

The official @deepseek-ai/dsh-mcp-client mounts one static plugin row per MCP server in cordis.yml, so adding or removing a server means editing the composition file and restarting the host. dsh-mcp-manager fills that management gap: add, remove, enable, disable, and reconnect MCP servers right inside a conversation with zero host restarts, aligning the MCP toolchain's operational experience with dynamic service patterns.

Three capability layers. A persisted JSON registry (default $DSH_HOME/mcp-manage/servers.json) stores server addresses, environment variables, and headers on disk; on host restart every registered server reconnects automatically. Values support ${VAR} expansion from the host process environment, writes are atomic (tmp file plus rename), and manual edits take effect on the next reload or restart. Six mcp_manage_* agent tools (list, add, remove, set_enabled, reload, tools) cover the full server lifecycle mid-session. Bridged tools follow the exact mcp____ naming contract and the same deterministic hash-suffix truncation algorithm as dsh-mcp-client, so the model cannot distinguish a dynamically managed server from a static row. The Settings to MCP Servers page in the Web UI provides visual add and edit for stdio or streamable-http endpoints, live connection status with actionable error messages, per-server and bulk reconnect, and a browsable tool listing; edits use a remove-then-add strategy so the server reconnects and the tool list refreshes immediately after saving.

Safety and stability mirror the official client throughout. The stdio subprocess environment is scrubbed of credential-shaped variables (KEY, TOKEN, SECRET, PASSWORD, case-insensitive) and all DSH_-prefixed variables before an explicit env block is merged back in. Disconnections trigger exponential-backoff reconnection starting at 500 ms and capping at 30 s; a stable connection resets the budget, and continuous failures beyond the max attempt ceiling stop retrying and unregister tools until mcp_manage_reload is issued. Namespace conflicts with other plugin rows fast-fail with a clear error message without consuming the reconnect budget. MCP tools list_changed notifications drive a generation-swap of the tool set, keeping the previous generation on failure and rolling back on conflict. Every registration, connection, and timer is scoped to the plugin effect, so卸载 (removal, including HMR) cleans up atomically.

Built for developers already running MCP toolchains on DSH who want to eliminate the edit-and-restart loop, and for platform or ops teams that share server configurations without hard-coding secrets into the profile manifest.