# DSH MCP Console (dsh-mcp-console) [简体中文](README.md) · **English** **Manage MCP servers in DeepSeek Harness: see their status, switch them on and off, let AI install one from a repository, and discover local software that can be connected.**

## Features - **MCP list** — an "MCP" sidebar entry lists every MCP server of the current profile: name, connection status, transport, command or URL, environment variable names, tools, and repository link. - **Enable / disable** — one switch per server, written through DSH's official Plugin Manager and applied immediately. - **AI quick add** — enter where the server comes from: a code repository, an npm or PyPI package, a Docker image, a launch command from a README, or any web page. A new session opens in the default workspace and the AI works out how to run it, installs it, writes the configuration, and connects it. Every install step asks for your approval. - **Add manually** — when you already know how it runs, enter a local command (`npx -y `, `uvx `, `docker run -i --rm `) or a remote endpoint, without the AI. Secrets can be read from environment variables instead of being stored. - **Search the MCP Registry** — search the official registry; an option fills Add manually with its command, environment variables, and headers. - **Test, edit, export** — Add manually tries the server (MCP handshake and tool list) before installing it, warns about missing environment variables and runtimes, and can go through mcp-remote for SSE-only or OAuth servers; servers in the plugin's bundle directory can be edited; all servers can be exported as `mcpServers` JSON without secrets. - **Discover local services** — scan installed applications, background services, listening ports, Docker containers, and MCP servers already configured in Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, and Codex; then let an AI session research which of them have MCP servers and ask which to connect. - **No secrets shown** — only the names of environment variables and headers are shown, credential-looking arguments are masked, and the scan inventory carries no file paths. ## Install In the desktop app, open **Plugins → Add plugin**, choose the official registry, and enter `dsh-mcp-console`. Within 24 hours of a release, install `https://github.com/CoderKuo/dsh-mcp-console/releases/latest/download/dsh-mcp-console.tgz` instead. ## How it works - **Status**: `@deepseek-ai/dsh-mcp-client` publishes no connection state, so it is inferred: *connected* = running with `mcp____*` tools; *connecting* = starting, or running without tools yet (also a failed first connection or a server without tools — see the DSH log); *failed* = the row failed to load (only with `failOnStartupError: true`); *disabled* = switched off. - **Enable/disable** calls the Plugin Manager's `setPluginEnabled`, which writes `disabled` into the profile patch. Rows provided by a bundle or another layer are read-only and say why. - **AI quick add** classifies the source (repository, npm, PyPI, package name, Docker image, or web page / endpoint), opens a Creator-mode session (agent preset `cordis`, which carries the `plugin_manager` tool), and follows the documented procedure: read the source, run it with the launcher that fits (`npx -y`, `uvx`, `docker run -i --rm`, or `streamable-http`) once, write a configuration-only bundle under `~/.dsh/mcp-bundles/`, install it with `plugin_manager install_bundle`, and verify the tools. Secrets are referenced as `!!js process.env.NAME`, never written literally. - **Add manually** builds the `@deepseek-ai/dsh-mcp-client` config from the form (the command line is split like a shell line but never run through a shell), writes a configuration-only bundle under the bundle directory, and installs it with the Plugin Manager's `installBundle`. A value marked "from environment" is written as `!!js process.env.NAME` (or a template such as `Bearer ${API_TOKEN}`); set the variable and fully restart DSH. - **Discovery** runs locally through `ctx.subprocess` (one PowerShell query on Windows; `ps`, `lsof`/`ss`, `systemctl`, `/Applications` elsewhere). Only when you click "Analyze with AI" is the inventory sent to your model as a message. ## Settings Row id `mcp-console` in the profile's `cordis.patch.yml`: `agentPreset` (default `cordis`), `bundleDirectory` (default `/mcp-bundles`), `scanTimeoutMs` (default `20000`), `repositories` (edited from the page). ## Development ```sh pnpm install pnpm run check ``` License: [MIT](LICENSE)