@looptech-ai/understand-quickly-mcp

A thin Model Context Protocol server that exposes the understand-quickly registry to any MCP client (Claude Desktop, Codex, Cursor, etc.).

Status: stub-quality. It works end-to-end but is intentionally minimal — no streaming, no embeddings, no auth.

What it does

It wraps the public registry.json and exposes four tools:

The registry response is cached in-memory for 60 seconds. stats.json uses an identical 60-second TTL cache.

find_graph_for_repo

Accepts either an id (the registry id, owner/repo) or a github_url. The URL parser tolerates:

• https://github.com/owner/repo
• https://github.com/owner/repo.git
• https://github.com/owner/repo/ (trailing slash)
• https://github.com/owner/repo/tree/main/... (branch / sub-path)
• git@github.com:owner/repo.git

When the entry is found, the response includes last_synced, last_sha, source_sha, head_sha, commits_behind, and a pretty drift_summary (e.g. "behind by 17 commits") when those fields are present in the registry.

If the entry is not found, the response is { found: false, suggestions: [...] } with up to 5 fuzzy-matched ids (Levenshtein distance ≤ 3 against the lowercased id).

search_concepts

By default — that is, when id is not provided — search_concepts reads the precomputed stats.json aggregate (a single, cached GET) and returns matching concept terms with their entry counts and up to 3 sample registry ids. This replaces the previous behaviour, which fanned out up to 5 graph fetches at request time.

When id is provided, it falls back to the legacy single-graph node search (substring match against id / label / name). When stats.json is unavailable (404 or schema mismatch), it falls back to the capped cross-graph fan-out for backward compatibility.

The source field on the response indicates which mode served the request: "stats", "graph", or "fanout".

Install

cd mcp
npm install
npm run build   # compiles TypeScript -> dist/
npm test        # runs node:test across registry/cache and tool logic

Node 20+ is required (uses the global fetch).

Run locally

For development:

npm run dev

For a built binary:

npm start

The server speaks stdio JSON-RPC. It will not respond to keystrokes — point an MCP client at it.

Register with Claude Desktop

Add the following to Claude Desktop's claude_desktop_config.json (the path is ~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "understand-quickly": {
      "command": "npx",
      "args": [
        "tsx",
        "/absolute/path/to/understand-quickly/mcp/src/index.ts"
      ],
      "env": {
        "UNDERSTAND_QUICKLY_REGISTRY": "https://looptech-ai.github.io/understand-quickly/registry.json"
      }
    }
  }
}

Replace /absolute/path/to/... with the actual path to your checkout. Restart Claude Desktop after saving.

If you would rather run the compiled output, swap to:

{
  "mcpServers": {
    "understand-quickly": {
      "command": "node",
      "args": ["/absolute/path/to/understand-quickly/mcp/dist/index.js"]
    }
  }
}

Environment variables

Current limitations

• In-memory cache only. Every server process refetches once a minute. No cross-process or on-disk cache.
• Cross-graph fan-out is only a fallback. When search_concepts falls back (no stats.json), it scans only the first 5 status: ok entries sequentially.
• Substring search is dumb. No fuzzy matching, no ranking, no embeddings.
• No streaming or progress reporting. Tools block until the upstream responds.
• Best-effort node enumeration. The single-graph fallback assumes the graph has a nodes / entities / concepts / items array; otherwise it walks top-level array values.
• No retries or backoff on upstream graph_url fetch failures — failed fetches return an empty result for that entry instead of erroring out.

These are all acceptable for an MVP. If you need more, open an issue.