@guowenzhang/dsh-mcp-manager
DeepSeek Harness 的 MCP 服务器管理：编写组合行，选择已允许的服务器何时加载，并筛选会话可调用的工具
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @guowenzhang/dsh-mcp-manager@0.2.0
Mcp Manager 以 @guowenzhang/dsh-mcp-manager 发布，当前版本为 0.2.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
中文 | English
DeepSeek Harness (dsh) is the open-source agent harness from DeepSeek AI, where nearly every capability is a plugin on Cordis. It is in developer preview and iterating fast, so expect compatibility-breaking changes (docs, 0.1.7-alpha.*); this plugin is a standalone third-party package that resolves @deepseek-ai/* from the running host.
MCP servers were hand-written composition rows whose whole tool set always sat in context, and an existing Claude Code setup had to be retyped; this plugin manages the rows from the settings page, loads a server on demand with only the tools you pick, and imports your Claude MCP configuration in one click.
The three loading modes and the configured servers, each row carrying its plane, live status, Edit, and enable switch.
A server's JSON configuration, and the tools it publishes: every one is ticked by default, and an unticked method never reaches the model.
Import Claude MCP configuration
Every MCP server found in the Claude Code configuration files; the ticked ones are imported as global rows.
npx @deepseek-ai/dsh plugin --profile web add @guowenzhang/dsh-mcp-manager
From the npm registry: https://www.npmjs.com/package/@guowenzhang/dsh-mcp-manager — restart the host afterwards; local checkouts, git sources and troubleshooting are in AGENTS.md.
The enable switch says whether a server may be used; the mode says when an allowed server enters context. Set it in 设置 → MCP 管理 → MCP 加载方式; the choice is stored in the mcp-manager settings namespace and applies from the next request on.
| Mode | Behavior |
|---|---|
Load all (eager) | Allowed servers mount at session start; their tools are always in the request |
Dynamic insert (dynamic, default) | They stay unmounted; mcp_load mounts one into the calling session — best tool binding, but the tool list changes once per load. A filtered row mounts only its visible tools |
Lazy (lazy) | mcp_load connects without registering anything and returns the tool schemas, called through the fixed mcp_call proxy — the tool list never changes, so the request-cache prefix is never invalidated |
Under dynamic / lazy the system prompt lists every loadable server as name — the description you wrote on its row, and the model calls mcp_load / mcp_unload by name. Names are always listed; descriptions are truncated to 80 characters under a 900-character budget. Load state is deliberately absent — reporting it would rewrite the system prompt on every mcp_load and invalidate the whole cache prefix.
Connections are per-session: a repeated mcp_load reuses one, different sessions each get their own, and a session that ends closes what it opened; eager shares one standing instance instead. Toggling a row shows a brief starting / stopping state while the child process comes up.
Settings → MCP 管理 → a row's Edit → the Tools tab lists every method the server publishes, all checked. Unchecking one hides it: it never enters context, and calling it is refused. Filtering does not change the cost model — the loading mode decides that.
For wildcards, write the rules yourself in the mcp-manager settings under tools, keyed by the row key (preset:<preset id>:<serverName>):
| Form | Meaning |
|---|---|
create_workitem | Allow list: one entry without ! keeps only matching tools |
!delete_* | Deny list: when every entry starts with !, matching tools are hidden |
*, ? | Wildcards: * matches any run of characters, ? matches exactly one |
Rules are read at the next mcp_load; an already-loaded server keeps the tools it was admitted with. eager ignores filters, and an unparsable rule set hides nothing.
设置 → MCP 管理 → 导入 Claude 配置 reads the files Claude Code writes and offers every server it finds as a checklist; ticked servers are imported as global rows.
| Source | File |
|---|---|
| Claude Code user config | ~/.claude.json → mcpServers |
| Per-directory scope inside it | ~/.claude.json → projects["<working directory>"].mcpServers |
| Claude Code settings | ~/.claude/settings.json, settings.local.json |
| Project scope | <project root>/.mcp.json |
Nothing is modified — the scan only reads — and each server is imported on its own, so one failure does not stop the rest. env / headers come along so the server can connect, but the dialog shows only those keys' names.
npx -y … / uvx …, usually 1–3 seconds). The UI never blocks; installing the server as a direct executable shortens this noticeably..mcp.json only. Cursor, Cline, Roo, and VS Code configuration files are not scanned, and an import always targets the global plane; move a row into a preset afterwards if you want it on-demand.dynamic gets no server instructions and no resource tools. The harness's mcp-client provides both, and this row registers through the plugin's own carrier instead. The tools themselves — argument binding, results, image projection — match a native mount.Apache License 2.0 — see LICENSE. This project includes MIT-licensed portions derived from DeepSeek Harness; see NOTICE.