Live now · free & read-only
Skillselion MCP
Community-proven skills for your coding agent, loaded mid-task.
The Skillselion MCP is a Model Context Protocol server that gives your coding agent on-demand access to 80,000+ skills, MCP servers and marketplaces, ranked by real installs. It finds the right one for the task and loads its real instructions mid-task - no install. Works with Claude Code, Claude Desktop, Cursor, Codex and any MCP client.
- No install
- Skills · MCP servers · marketplaces
- Ranked to your repo
What is the Skillselion MCP?
It turns the Skillselion directory into a live capability your agent can call. Instead of reciting best practices from memory, your agent searches the catalog and pulls a skill’s real instructions into the current task. Four tools:
- load_skill
- Pull a skill’s real SKILL.md and its bundled scripts and references into the task. No install.
- synthesize_skills
- Merge the key rules of the top matching skills into one consensus playbook, every rule attributed to its source skill.
- search_skillselion
- Search by task or keyword, across skills, MCP servers and marketplaces.
- top_skillselion
- The leaderboard, ranked by real installs and GitHub stars.
How does it work?
Find the skill. Load it. Follow it. In the same task, with nothing installed.
- 01Search the catalog by task or keyword.
- 02Load the chosen skill’s real instructions into a temp folder.
- 03Follow the proven recipe, mid-task.
synthesize_skills - a consensus playbook merged from 5 community skills, every rule attributed - then load_skill to apply the top skill and rewrite the CSS live. It closes with Claude’s own words: “this is the field’s consensus, not my opinion - that’s the part I can’t fake.”How does it find the right skill?
It blends keyword search with semantic search powered by OpenAI’s text-embedding-3-small (1536-dimension) embeddings stored in Postgres with pgvector, then ranks candidates by real installs, GitHub stars and editorial curation, fitted to your repo’s stack. So a query like “listen to database row changes in realtime” surfaces the right tools even when they share no keyword with your request.
How is it different from find-skills?
find-skills (by Vercel) installs a skill into your agent. The Skillselion MCP loads the right tool into the current task with no install, and covers MCP servers and marketplaces too, not just skills. Both rank by community signal and both read skills.sh, so the difference is the model, not the data.
| Skillselion MCP | find-skills | |
|---|---|---|
| How you use it | Loads the skill into the current task (temp folder) | Installs the skill into your agent config |
| Coverage | Skills + MCP servers + marketplaces | Skills only |
| Ranking | Ranked to your repo + installs, stars, editorial | Sorted by install count |
| Footprint | Nothing installed; gone after the task | Persistent install on your machine |
| Form factor | MCP server (any MCP client) | A skill that drives the npx skills CLI |
How do I install it?
It’s live and free on npm, and the server is open source (MIT) on GitHub - issues and pull requests welcome. One command in Claude Code:
$ claude mcp add skillselion -- npx -y skillselion-mcpCursor, Codex or Claude Desktop - add to your MCP config:
{
"mcpServers": {
"skillselion": { "command": "npx", "args": ["-y", "skillselion-mcp"] }
}
}Read-only on your machine: it never installs to or modifies your repo, and your code never leaves your machine. It sends only your search query (set DO_NOT_TRACK=1 to disable). No auth, no secrets.
FAQ
Is the Skillselion MCP free?
Yes. It is free and read-only, and works with any MCP client - Claude Code, Claude Desktop, Cursor and Codex.
Do I have to install each skill?
No. load_skill pulls a skill’s real instructions into the current task from a temporary folder. Nothing is added to your project or your agent config.
Does it cover MCP servers and marketplaces, or only skills?
All three. You can search skills, MCP servers and plugin marketplaces; load_skill focuses on loading skills into the task.
Does it use semantic search?
Yes. It blends keyword search with embedding-based semantic search (OpenAI text-embedding-3-small with pgvector), so it finds the right tool even when it shares no keyword with your request.
What data does it send?
Only your search query - scrubbed of emails, tokens and file paths and capped at 200 characters - plus whether it matched and which skill loaded, so the catalog learns what to add next. No identity, no IP, and never your code, your files, or the context you pass. Set DO_NOT_TRACK=1 (or SK_NO_DEMAND=1) to disable it.
Will it bloat my agent context?
No. load_skill uses progressive disclosure: a compact card plus the skill instructions, and for very large skills it clips at a section boundary and leaves the full copy on local disk to read on demand. You get the recipe without dumping thousands of tokens into context.
How is it different from find-skills?
find-skills installs a skill into your agent. The Skillselion MCP loads the right skill, MCP server or marketplace into the current task with no install, ranked to your repo.