dsh-mcp-manager

English · 中文

Durable MCP server manager for DeepSeek Harness (DSH) — a composed loader plugin (not a dynamic session plugin), so it survives DSH restarts and upgrades.

Built to the DSH plugin development standard (see the official docs: 第一个插件, 开发一个工具):

• object-form Cordis plugin (export default { name, inject, apply })

• required services declared in inject — the framework guarantees they are ready before apply runs, and reloads the plugin if one disappears (this is what keeps the plugin alive across DSH upgrades)

• agent-facing capabilities exposed as registered model tools via ctx.tools.register(defineTool(...)): mcp_manager_list, mcp_manager_set_enabled, mcp_manager_restart, mcp_manager_add

• UI-facing capability via a webServer exact route (POST /dsh-mcp-manager/api), consumed by the client half

Features:

• Settings → MCP 管理 page (added by the client half)

• Host half manages @deepseek-ai/dsh-mcp-client rows in the real patch files:

  • 项目级 → ~/.dsh/profiles/<profile>/cordis.patch.yml

  • 全局 → ~/.dsh/cordis.patch.yml

  • add / edit / enable / disable / restart / delete, live tool-count health, JSON export/import, per-file write lock

How it works

The loader entry is the single composition point for both halves — no changes to any shipped DSH package. The package itself is platform-neutral (pure JS; path separators detected at runtime), so it works on Windows, macOS and Linux. Building (npm run build) requires the devDependencies (typescript, @deepseek-ai/cordis, @deepseek-ai/dsh-tools, @types/node); the shipped/installed package needs none of them.

Related MCP server: Deepseek MCP Server

Install (any platform)

Quickest — one command, via npm (requires Node.js >= 18):

# Option A — one-shot with npx, nothing to install
npx -y @xxxyz/dsh-mcp-manager

# Option B — global npm install, then run the command any time
npm i -g @xxxyz/dsh-mcp-manager
dsh-mcp-manager                  # install the plugin
dsh-mcp-manager-uninstall        # uninstall the plugin
npm i -g @xxxyz/dsh-mcp-manager@latest   # upgrade

All flags pass through in both options: npx -y @xxxyz/dsh-mcp-manager --dsh-home /path/.dsh --profile web --repair --port 3080.

npm package name: @xxxyz/dsh-mcp-manager (the bare name dsh-mcp-manager is taken on npm by an unrelated package). The deployed plugin name is still dsh-mcp-manager — npm/npx is only the delivery channel; the installer copies the files to the same fixed locations below.

No-npm alternative — directly from GitHub:

npx -y github:xxxyz/DeepSeekHarness-MCP-Manager

Alternative — run the installer scripts from a source checkout (all installers share one cross-platform logic file, install.mjs). Install steps:

1. copy the package to <dshHome>/local-packages/dsh-mcp-manager (a true source kept outside node_modules, so DSH upgrades never touch it)

2. copy it into <dshHome>/profiles/node_modules/dsh-mcp-manager (a plain copy on purpose — a symlink would make Node ESM resolve the plugin's realpath where @deepseek-ai/dsh-tools cannot be found)

3. append the loader row to profiles/<profile>/cordis.patch.yml (idempotent; keeps the patch a valid top-level array)

Windows (PowerShell):

.\dsh-mcp-manager\install.ps1                     # default: ~/.dsh, web profile
# or: .\install.ps1 -DshHome D:\path\.dsh -Profile web

macOS / Linux:

./dsh-mcp-manager/install.sh                      # default: ~/.dsh, web profile
# or: ./install.sh --dsh-home /path/.dsh --profile web

Any platform (direct):

node dsh-mcp-manager/install.mjs [--dsh-home <path>] [--profile <name>] [--port <n>] [--repair] [--skip-patch]

Then restart DSH and open Settings → MCP 管理. The four mcp_manager_* tools become callable by the model after that restart.

Uninstall (any method):

dsh-mcp-manager-uninstall          # if installed via npm -g
# or: .\dsh-mcp-manager\uninstall.ps1 | ./uninstall.sh | node uninstall.mjs [--dsh-home <path>] [--profile <name>]

then restart DSH. Uninstall removes the deployed copy, the local-packages true source, and the loader row.

After a DSH upgrade: --repair

A DSH upgrade (or a broken HMR state) can leave the plugin's host half unloaded while the files are still in place. One command fixes it — re-copies the package from the source of record, bumps the loader row's config.version to force an HMR re-apply, then polls the API until it answers:

node dsh-mcp-manager/install.mjs --repair            # default ~/.dsh, web, port 3080
node dsh-mcp-manager/install.mjs --repair --port 3080
# PowerShell: .\install.ps1 -Repair -Port 3080     bash: ./install.sh --repair --port 3080

--repair returns success only when POST /dsh-mcp-manager/api answers {ok:true} again. If the API still does not answer after 30s, restart DSH once (the loader always re-imports fresh at boot).

API reference

All ops are POST /dsh-mcp-manager/api with {"op": "<op>", "args": {...}}, same-origin.

Model tools

Registered on the host with ctx.tools.register(defineTool(...)) (standard @deepseek-ai/dsh-tools):

Notes / limitations

• The HTTP route is unauthenticated on the local web server — fine for a local single-user machine; do not expose the DSH web port publicly.

• Managed rows carry # dsh-mcp-manager:server:<id> markers; the loader row is an insert block adding id: mcp-manager, name: dsh-mcp-manager — DSH's patch dialect treats a plain - id: row as an override of an existing entry (silently skipped when absent), so adding a plugin requires the insert form.

• If the running web page predates the install, a page refresh or DSH restart is needed for the client module (the boot graph is built at page load).