---

description: "MCP server management for DeepSeek Harness: author composition rows, choose when an allowed server loads, and filter which of its tools a session may call." kind: "plugin-readme"

@zhang-guo-wen/dsh-mcp-manager

中文 | English

What this plugin does

• MCP management. Add, edit, enable, and disable MCP servers from the settings page. Global rows and every preset's rows are listed with live status. A toggle shows a brief starting / stopping state because the child MCP process has to come up; the list never blocks on it.
• On-demand loading. A stopped server costs nothing. The model calls mcp_list, mcp_load, and mcp_unload to start a server for the calling session only; loaded tools never leak into another session. Connections are per-session, and a session that ends closes the connections it opened.
• MCP tool filters. The edit dialog lists the methods a server publishes with every one checked; an unchecked method never enters context — neither listed nor callable. Wildcards are available by writing mcp-manager.tools yourself.

MCP management was split out of dsh-claude-compat into its own plugin. The two do not depend on each other and install separately; with both installed, the settings page shows Claude 兼容 and MCP 管理 as independent sections.

Screenshots

MCP 管理 — every configured server, live

MCP 管理 — add or edit a server, including the tool picker

Install

The built lib/ is committed, so the repository installs and runs directly — no build step on your machine.

# over HTTPS
npx @deepseek-ai/dsh plugin --profile web add git+https://github.com/zhang-guo-wen/dsh-mcp-manager.git

# or over SSH
npx @deepseek-ai/dsh plugin --profile web add git+ssh://git@github.com/zhang-guo-wen/dsh-mcp-manager.git

Pin a release tag so a later work-in-progress commit on the default branch is not picked up:

npx @deepseek-ai/dsh plugin --profile web add "git+ssh://git@github.com/zhang-guo-wen/dsh-mcp-manager.git#v0.1.0"

To develop against a local checkout, install the directory: pnpm links it, so rebuilding lib/ takes effect on the next start without reinstalling.

npx @deepseek-ai/dsh plugin --profile web add /absolute/path/to/dsh-mcp-manager

Use

MCP loading modes

A row's enable switch and the loading mode answer different questions: the switch says whether a server may be used at all, the mode says when an allowed server enters context.

Mode	Behavior
Load all (eager)	Allowed servers mount at session start; their tools are always in the request
Dynamic insert (dynamic, default)	Allowed servers stay unmounted; mcp_load mounts one into the calling session, so its tools join the request — best tool binding, but the tool list changes once per load
Lazy (lazy)	Allowed servers stay unmounted; mcp_load connects over the MCP SDK without registering anything and returns the tool schemas, and the model calls them through the fixed mcp_call proxy — the tool list never changes, so the request-cache prefix is never invalidated

Set it in 设置 → Harness 兼容 → MCP 管理 → MCP 加载方式. The choice is stored in the user's mcp-manager settings namespace and applies from the next request on, in every session.

Processes and lifetime

Under dynamic / lazy, a server that has not been loaded starts no process at all — MCP is stopped when the session begins, until some mcp_load.

Once loaded, connections are per-session: a repeated mcp_load in one session reuses the same one, while different sessions each get their own (for stdio, one child process each); subagents and forked sessions count as separate sessions. A session that ends closes the connections it opened, with no mcp_unload required. eager is the opposite — the preset is a standing mount, so one shared instance serves every session.

MCP rows on the global plane (written directly into cordis.yml) are outside on-demand loading: they always start, as if permanently eager. Put a server in a preset to make it on-demand.

MCP tool filters

A server often publishes dozens of tools while a session uses a few. Once filtered, only the tools the rules admit are handed to the model when the server loads: a hidden tool is absent from the mcp_load result and mcp_call refuses to invoke it.

In the settings page: Settings → Harness 兼容 → MCP 管理 → a row's Edit → the Tools block at the bottom of the dialog.

Opening it connects to that server once, lists the methods it publishes, and checks every one of them. Unchecking a method disables it; the change applies on save. The block carries an enabled/total count, Load tools (re-read after editing the JSON), and All / None.

• Rules are only rewritten when the server actually answered; a failed connection leaves the stored rules untouched.
• Everything checked = no rules for that row, so every published tool stays visible (also the state of a new row).

For wildcards, write them yourself. Rules live in the mcp-manager settings namespace under tools, keyed by the row key (preset:<preset id>:<serverName>, the same key the description map uses):

Form	Meaning
create_workitem, get_workitem	Allow list: one entry without ! means only matching tools stay visible
!delete_*	Deny list: when every entry starts with !, matching tools are hidden and the rest stay
*, ?	Wildcards: * matches any run of characters, ? matches exactly one

What the dialog saves is exactly that deny list, so unchecking delete_workitem writes:

mcp-manager:
  tools:
    "preset:standard-yunxiao:alibaba-devops-mcp":
      - "!delete_workitem"

What to expect:

• Rules are read at load time. A committed change applies to the next mcp_load; an already-loaded server keeps the tools it was admitted with, and mcp_unload followed by mcp_load picks up the new rules.
• A filtered row always takes the proxy carrier (mcp_load lists, mcp_call invokes), even under the dynamic mode: native registration publishes every discovered tool and offers no way to hold some back.
• eager ignores filters, because that mode mounts the whole server through the harness's mcp-client. The plugin warns at startup when rules are configured for it.
• A malformed rule set hides nothing: an unparsable value filters nothing, so a typo never empties a server.

Configuration

Field	Default	Meaning
mcpLoading	dynamic	How allowed servers enter context by default; the settings-page choice overrides it

- name: '@zhang-guo-wen/dsh-mcp-manager'
  config:
    mcpLoading: lazy

Known limitations

• Enabling a server still waits on the child process (npx -y … / uvx …, usually 1–3 seconds). The UI never blocks; installing the server as a direct executable shortens this noticeably.
• Opening the edit dialog connects to that server once (to list its tools), with the same 1–3 second cost.
• Global-plane rows ignore the loading mode: they always mount.
• A preset's first mount starts and then kills each server once. On-demand loading works by unmounting rows at runtime, which cannot beat the child process's spawn, so the first session to use a preset after a host restart pays one short start-up.

Development

Build, Cordis/Typert plugin contract, the traps, and the MCP lifecycle details live in AGENTS.md.

npm run build      # host (tsdown) + client (rolldown ModuleLoader handoff)
npm run typecheck

License

Apache License 2.0 — see LICENSE. This project includes MIT-licensed portions derived from DeepSeek Harness; see NOTICE.