alone-tree/dsh-skill-mcp-manager ↗★ 6

dsh-skill-mcp-manager

能力库 (Capability) — one-stop visual management of DSH Skills & MCP: on-demand MCP loading, in-session hot reload, and SKILL/MCP descriptions viewable in the plugin. 适合需要动态加载、热重载MCP服务及精细化控制工具开关的开发者和高级用户。

패키지
dsh-skill-mcp-manager
호환성
미검증
Harness peer 범위
^0.1.0-rc.6 || ^0.2.0-rc.1
Cordis peer 범위
^4.0.1
버전
1.2.1
라이선스
MIT
최근 업데이트
2026. 10. 3.

같은 패키지 이름의 다른 저장소

설치

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:alone-tree/dsh-skill-mcp-manager

Usage

In the UI

Open Settings → Capability:

  • SKILL tab — every managed skill (bundle SKILL.md), sorted by path: model-visibility switch, open in system editor, cross-platform delete. Skills shipped with the deployment (under node_modules / app.asar) are read-only: view + open only.
  • MCP tab — every registered server with tier badges and tool counts: switch tier, inspect details (command / env / headers / tools), enable or disable individual tools (blacklist; newly discovered tools default to enabled), mask or reveal secrets, peek descriptions, refresh the shared tool snapshot, delete entries. Snapshot refresh trial-connects and updates the registry, then closes; it does not leave a live instance. Disabled tools are hidden from the AI and rejected again at the execution boundary. A setting row adjusts the catalog's tool-description truncation (default 150 chars).

Model-facing tools

The plugin provides three small, fixed tools to manage all on-demand MCPs.

  • mcp_register — add or modify an MCP entry. Trial-connects (30s) + lists tools, then persists; the server's own description / version / title / website / instructions are captured automatically. Reuse it to change tier / parameters / notes; only connection-contract changes re-connect.

    mcp_register { name, tier?, transport, command?, args?, env?, cwd?, url?, headers?, notes? }
    

    tier ∈ eager | on-demand | disabled (default on-demand). notes is user-maintained and is not overwritten by developer MCP updates.

  • mcp_load { name, peek? } — load / hot-reload a server instance for the current session, returning full tool definitions + server-declared metadata; other sessions are unaffected. peek: true only reads the snapshot — no connect, no disconnect — handy when the AI forgot a tool's parameters and wants a quick peek without interrupting the MCP's live process. Especially friendly for stateful MCPs like browser automation.

  • mcp_call { name, tool, args? } — invoke an on-demand tool (must mcp_load first) through the current session's loaded instance. Structured args only — never shell text, never a temp file. Malformed shapes (missing/non-string tool, non-object args) are rejected with the correct shape spelled out, never auto-unwrapped; failed calls (protocol errors and server isError results) echo back called MCP tool: / plus the exact arguments sent so the model can self-correct.