Skillselion MCP

What skillselion/skillselion-mcp MCP server does

skillselion/skillselion-mcp MCP server exposes a read-only catalog of community skills, MCP servers, and marketplaces from Skillselion. Catalog results are ranked using reported installation counts and GitHub stars. The server is intended for coding agents that need to discover relevant instructions during a task rather than relying only on their built-in knowledge.

The main workflow is to ask load_skill for a task. It finds a relevant entry, places the skill's actual SKILL.md content into the current context, and writes bundled scripts or reference files to disk. The response also includes ranked alternatives that the agent can choose instead. If no sufficiently relevant match is found, the tool can decline to load one rather than adding unrelated context.

How it works

The package runs as a local MCP process started with npx -y skillselion-mcp. Its tools query the Skillselion catalog and, when needed, retrieve skill content from GitHub. search_skillselion accepts a keyword or task and can limit results to skills, MCP servers, or marketplaces. top_skillselion returns popular skills and highly starred MCP servers.

synthesize_skills combines the key rules from roughly five matching skills into a deduplicated digest with source provenance. The number of sources and several output limits can be adjusted through environment variables. The server does not require authentication. It sends a scrubbed search query and the matching skill identifier as an anonymous demand signal unless tracking is disabled with DO_NOT_TRACK=1; it does not send the code, files, or task context described in the README.

Setup and configuration

For Claude Code, register the process with:

claude mcp add skillselion --scope user -- npx -y skillselion-mcp

Other MCP clients can use a stdio entry equivalent to:

{
  "mcpServers": {
    "skillselion": {
      "command": "npx",
      "args": ["-y", "skillselion-mcp"]
    }
  }
}

No build step, account, API key, or required environment variable is documented. An optional read-only GITHUB_TOKEN raises the GitHub request limit for multi-file skill loads. SK_INLINE_BUDGET limits the in-context SKILL.md portion, while SK_SYNTH_N, SK_SYNTH_PER_SKILL, and SK_SYNTH_BUDGET control synthesis size. Set DO_NOT_TRACK=1 to disable the anonymous demand signal.

The package also includes an optional setup command:

npx -y skillselion-mcp setup
npx -y skillselion-mcp setup --yes

It can register the server and add a Claude Code SessionStart hook. Packs include popular, frontend, ai-agents, media, backend, devops, quality, and automation. Options can select packs, set per-pack counts, adapt selection to the current repository, or inspect Claude Code and Codex history once.

Tools and capabilities

• load_skill: Finds a matching skill and materializes its instructions and bundled files.
• synthesize_skills: Produces a provenance-tagged digest from multiple matching skills.
• search_skillselion: Searches the catalog and filters by listing type.
• top_skillselion: Shows current popular skills and top-starred MCP servers.

Limitations and notes

The catalog is read-only; this MCP server does not install or modify catalog entries. Full multi-file loads can be affected by GitHub's unauthenticated request limit, in which case a read-only token may be needed. Some clients expose MCP tools lazily, so the agent may need to search for the skillselion tools before calling them. The optional setup hook is specific to Claude Code, while the MCP process itself can be configured in other clients that support stdio MCP servers.