lilyblessing/dsh-mcp-skill-panel
MCP and skill management panel: MCP server/Skill real-time start and stop release context (number of backfill directory tools in disabled state); optional AI middle layer (mcp_search/mcp_call) filters visibility according to server status, keep-alive enabled + idle 30s recycling
MCP & Skill Manager turns the MCP server list and Skill catalog into an actionable settings panel: one toggle per entry, where disabling an MCP server disposes its loader entry (closing the connection and unregistering its mcp<server>* tools so their schema tokens are freed immediately) and enabling reconnects without a restart; Skill toggles inject or remove disable-model-invocation: true in SKILL.md frontmatter. It adds an optional AI middle layer (autoManage) that hides disabled servers from the model and calls them on demand through dsh_mcp_search / dsh_mcp_call with idle reaping, keeps user-enabled servers always visible, supports a quick migration that converts another harness's mcpServers JSON (Claude Code/Codex style) with inferred transport and ${VAR} interpolation, and carries tool-level and bulk tool toggles with a three-state toolNames contract documented in the README.
Install
dsh plugin --profile web add "github:lilyblessing/dsh-mcp-skill-panel#main"The README uses the git spec pinned to #main; it also references an npm package named dsh-mcp-skill-panel (registry re-verified 2026-09-30: 0.6.0, repository field points back at github.com/lilyblessing/dsh-mcp-skill-panel, MIT) and warns that npm releases may lag the repository, noting that 0.5.4 / 0.5.5 were published and later unpublished, so the git source has the latest code. A from-source install runs npm install --legacy-peer-deps --ignore-scripts once.
Compatibility
DeepSeek Harness Web settings page. State persists to ~/.dsh/dsh-mcp-skill-panel/state.json with catalog snapshots in catalog.json. Project-level MCP is read from <workspace>/.dsh/mcps/**/mcp.json and is visible only to sessions of that workspace (filtered by session cwd), with file changes hot-reloading.
Details
- Repo: lilyblessing/dsh-mcp-skill-panel
- Category: Skill Packs
- Stars: 4
- Version: git spec github:lilyblessing/dsh-mcp-skill-panel#main (npm dsh-mcp-skill-panel 0.6.0 also published, registry re-verified 2026-09-30); MIT
- Last push: 2026-09-16
- First seen: 2026-08-15
Recent updates
The README documents real-time MCP/Skill toggling without restart, state persistence into the preset composition file, the optional autoManage middle layer with on-demand calls, project-scoped MCP that hot-reloads, and the tool-level bulk endpoint's three-state toolNames semantics.
FAQ
- How do I install MCP & Skill Manager?
- Run dsh plugin --profile web add "github:lilyblessing/dsh-mcp-skill-panel#main". The README notes the npm package dsh-mcp-skill-panel exists but may lag the repository, so the git source is the latest code.
- Does disabling an MCP server need a restart?
- No. Per the README disabling disposes the loader entry, closes the connection and unregisters the server's mcp<server>* tools immediately so their schema tokens are freed; enabling reconnects and restores the tools without a restart.
- What does the optional AI middle layer do?
- With autoManage on, disabled MCP servers are hidden from the model and used on demand via dsh_mcp_search / dsh_mcp_call, while servers you enabled stay visible; AI-temporarily-enabled servers are auto-reaped after idle and never override your own toggles.
Alternatives
Hanihahaha/dsh-skill-mcp-manager · Js2Hou/dsh-mcp-manager · Fishquito7/dsh-skill-mcp-panel