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/ /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

Piece	File	Role
Host plugin source	src/index.ts	TypeScript source (the documented plugin shape); build with npm run build (tsc)
Host plugin (compiled)	lib/index.js (main)	Cordis object-form plugin: patch-file CRUD + 4 model tools + webServer exact route /dsh-mcp-manager/api (JSON {op, args} → {ok, ...})
Client bundle	lib/client.js (exports["./client"] + dsh.client)	Browser module: registers the settings page; calls the host via fetch('/dsh-mcp-manager/api')
Loader row	added to the profile's cordis.patch.yml	composes the host entry; the client-modules service scans enabled entries and serves the client bundle

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.

Install (any platform)

All installers share one cross-platform logic file (install.mjs). Install steps:

1. copy the package to /local-packages/dsh-mcp-manager (a true source kept outside node_modules, so DSH upgrades never touch it)
2. copy it into /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/ /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 
] [--profile ] [--port ] [--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 (same three ways):

.\dsh-mcp-manager\uninstall.ps1      # or: ./uninstall.sh   or: node uninstall.mjs [--dsh-home 
] [--profile ]

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": "", "args": {...}}, same-origin.

op	args	result
mcpm-list	{}	{ok, rows[], paths, errors[]}
mcpm-add	`{serverName, transport, url	command, args?, headers?, env?, level, enabled?}`
mcpm-edit	{id, level, ...fields}	{ok}
mcpm-set-enabled	{id, level, enabled}	{ok}
mcpm-restart	{id, level}	{ok}
mcpm-remove	{id, level}	{ok}
mcpm-export	{}	{ok, json, savedTo}
mcpm-import	{json}	{ok, added[], skipped[]}

Model tools

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

tool	description
mcp_manager_list	list all configured MCP servers (level, enabled state, live loader status, tool count)
mcp_manager_set_enabled	enable / disable one server (id, level, enabled)
mcp_manager_restart	restart one server (id, level)
mcp_manager_add	add a server (serverName, transport, url

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: 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).