MCP server

MCP (Model Context Protocol) is the open standard that lets AI agents and editors call external tools. Connect this server and your agent can search looprepo's vetted loop catalog, grade any loop prompt A–F, and generate safe-loop harnesses — no auth, no API key, rate limit 60 req/min per IP.

One-click install

Endpoint: https://mcp.looprepo.dev/mcp

Add to Claude →Add to Cursor →Add to VS Code →VS Code Insiders →

The Claude button opens Customize → Connectors — click “+ Add custom connector” and paste the endpoint above (Claude has no one-click MCP install). Cursor opens directly; the VS Code buttons go through vscode.dev’s install redirect. Everything else below is copy-paste.

Claude Code

claude mcp add --transport http looprepo https://mcp.looprepo.dev/mcp

Claude Desktop

Customize → Connectors → “+ Add custom connector” with the endpoint URL above — or add this to claude_desktop_config.json (uses the mcp-remote bridge; needs Node installed):

{
  "mcpServers": {
    "looprepo": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.looprepo.dev/mcp"]
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "looprepo": { "serverUrl": "https://mcp.looprepo.dev/mcp" }
  }
}

Any other MCP client

Generic Streamable HTTP config (Codex, Cline, Zed, custom SDK clients, …):

{
  "mcpServers": {
    "looprepo": { "type": "http", "url": "https://mcp.looprepo.dev/mcp" }
  }
}

Tools

loops_search — search the catalog by text, category, loop type, safety grade, or provenance — sorted by top/new/copies, cursor-paginated
loops_get — fetch one loop’s full prompt, usage notes, attribution, and stats by slug
loops_categories — list categories and loop types with live published-loop counts
loops_grade — grade any loop prompt A–F with the same deterministic, default-fail Safety Score that gates this directory — no LLM
loops_generate_harness — generate a complete safe-loop harness bundle (requires an email; 30/day per email, plus a per-IP cap)

Access & rate limits

No auth, no API key, no signup — all read tools are open.
60 requests/min per IP across the API (429 + Retry-After when exceeded).
Request bodies capped at 32KB (413 above that).
generate_harness additionally: 30 builds/day per email, per-IP cap.

Troubleshooting

405 Method Not Allowed. You hit an old deploy or sent a GET. The transport is POST-only JSON-RPC at mcp.looprepo.dev/mcp. If a client saw 405s before 2026-07-06, retry — the fix is deployed.

403 from Cloudflare. Cloudflare Bot Fight Mode used to block Python-based MCP clients by user-agent. That is resolved. If you still see a 403 with a cf-ray header, tell us — it never reached the server.

“It returns HTML”. looprepo.dev/mcp (this page) is documentation. The MCP endpoint your client connects to is https://mcp.looprepo.dev/mcp — a JSON-RPC transport, not a web page.

Endpoints

https://mcp.looprepo.dev/mcp

MCP transport (Streamable HTTP, JSON-RPC 2.0) — POST only

https://mcp.looprepo.dev/health

health check

Prefer plain HTTP? The catalog is also directly fetchable — see the agents page for catalog.json, catalog.txt and llms.txt. Want to check a loop before running it? Grade it →