Connect from your agent.
SkillsRegistry runs as a public Model Context Protocol server. Five read-only tools cover the full discovery surface — semantic search, single-skill lookup, leaderboards, trust breakdowns, composition resolution — backed by the same engine the web search uses.
No SDK, no auth flow, no OAuth — drop a JSON config into your MCP client and call the tools.
Install
Add to your agent
Streamable HTTP transport at https://api.skillsregistry.net/mcp. No auth for read tools. Discovery: .well-known/mcp.json.
One command in your shell — Claude Code wires it up and verifies the connection. Run /mcp in any session to confirm.
claude mcp add --transport http --scope user skillsregistry https://api.skillsregistry.net/mcp --scope user for --scope project to commit it to .mcp.json. v1 tool surface
-
search_skillsargs: query (required), limit, appetite, tags, category, runtimeEnv, visibility, portableSemantic search across the whole catalog (live count on the homepage heartbeat). Returns confidence-tiered results (T1/T2/T3) with trust signals.
-
get_skillargs: slug (required), versionFetch a single skill by slug — full manifest, trust breakdown, source provenance, mcpUrl.
-
list_leaderboardargs: kind ('trust' | 'trending' | 'agents' | 'composed' | 'forked') (required), limit, skillType, category, ecosystemTop-N skills by leaderboard kind. Human and agent signal channels stay separate.
-
get_trust_breakdownargs: slug (required)Circle-IR trust slice — overall 0-100 score, six dimensions (security, supply chain, quality, reliability, compliance, provenance), tier (verified / passing / advisory / failing / blocked), findings and content-safety verdict. Breakdown is null for shallow-scanned skills, which is most of the catalog.
-
resolve_compositionargs: slug (required)Resolve a composition slug into its ordered constituent skills (DAG).
-
searchargs: query (required)ChatGPT-connector shape of search_skills: returns {results:[{id,title,url}]} (id = slug) for chat and deep research.
-
fetchargs: id (required — a slug from search)ChatGPT-connector shape of get_skill: returns {id,title,text,url,metadata} with a Markdown rendering of the skill.
Tested clients
- Claude Code
- One command: claude mcp add --transport http --scope user skillsregistry https://api.skillsregistry.net/mcp.
- Claude Desktop
- Settings → Connectors → Add custom connector → paste the endpoint URL. No authentication.
- ChatGPT
- Settings → Connectors → Create (developer mode) → MCP server URL. Chat and deep research use the connector search / fetch tools.
- Codex CLI
- ~/.codex/config.toml: [mcp_servers.skillsregistry] url = "https://api.skillsregistry.net/mcp".
- Cursor
- .cursor/mcp.json with "url", or the one-click install link in the Clients tab above.
- VS Code / GitHub Copilot
- .vscode/mcp.json: servers.skillsregistry = { type: "http", url }.
- Gemini CLI
- ~/.gemini/settings.json: mcpServers.skillsregistry.httpUrl.
- Windsurf
- ~/.codeium/windsurf/mcp_config.json: mcpServers.skillsregistry.serverUrl.
- Cline
- MCP settings: mcpServers.skillsregistry = { type: "streamableHttp", url }.
- Continue
- .continue/config.yaml: mcpServers entry with type streamable-http.
- Goose
- goose configure → Add Extension → Remote Extension (Streaming HTTP) → paste the URL.
- Anything else
- Any client speaking MCP 2025-06-18, 2025-03-26 or 2024-11-05 over Streamable HTTP works — that is what we serve.
Any client speaking MCP 2025-06-18 over Streamable HTTP works. The server negotiates protocol version on initialize.
Private overlays
Pass X-Tenant-Id: your-tenant on the HTTP request to union your private/unlisted skills with the public catalog. Missing or spoofed headers degrade to public-only — v1 treats it as an advisory scope hint, not a security boundary (OAuth 2.1 + RFC 8707 lands with v2 write tools).
Static JSON describing transport, tools, and auth posture. RFC 8615 well-known path.
Same data, OpenAPI-shaped. Use REST for browser apps, MCP for agents.