source-pack-mcp

An MCP server that produces a structured research source pack for a topic or question. Instead of an agent rummaging through search results and raw pages, it returns verified-at-source facts, quotes, key numbers, dates, primary links, and a coverage map of what's been said — as clean structured data.

What a source pack looks like

{
  "pack_id": "uuid",
  "query": "string",
  "created_at": "ISO timestamp",
  "sources": [
    {
      "url": "string",
      "title": "string",
      "domain": "string",
      "retrieved_at": "ISO timestamp",
      "facts": [{"fact": "string", "confidence": "high|medium|low"}],
      "quotes": [{"text": "string", "context": "string"}],
      "numbers": [{"value": "string", "unit": "string", "context": "string"}],
      "dates": [{"date": "string", "event": "string"}],
      "primary_links": ["string"]
    }
  ],
  "coverage_map": [
    {"claim": "string", "mentioned_by": ["domain"], "contested_by": ["domain"], "status": "consistent|contested|isolated"}
  ],
  "limitations": ["string"]
}

Facts are extracted from actually-fetched sources, with the URL attached and a confidence label. The server never invents facts and never claims verification it didn't perform. coverage_map compares claims across sources: consistent, contested, or isolated.

Tools

Tool	Purpose
pack_build	Build a source pack for a query (max_sources, search_depth).
pack_add_source	Add a specific URL to an existing pack for extraction.
pack_get	Retrieve a previously built pack by id.
pack_list	List recent packs (id, query, created_at, source count).
pack_search	Keyword search across previously built packs.

Features

• HTML → clean text via real DOM extraction (Mozilla Readability + Turndown), never regex parsing
• Disk cache with TTL (default 1 hour)
• robots.txt respect, rate-limited fetching (~1 req/sec)
• Packs persisted as JSON in a configurable data dir (.packs/)
• Streamable HTTP (default 127.0.0.1:3225) + stdio transports
• No API keys required, no auth needed for public content, no personal data

Quick start

Requirements: Node.js 22+

npm ci
npm run build
PORT=3225 node dist/src/http.js

Configure with env vars (see .env.example):

Var	Default	Purpose
SITE_BASE_URL / MAX_SOURCES_PER_PACK	8	Max sources per pack
FETCH_TIMEOUT_MS	15000	Per-request fetch timeout
FETCH_MIN_INTERVAL_MS	1000	Rate limit between fetches
SEARCH_API_URL	unset	Optional JSON search API; falls back to HTML search
CACHE_TTL_SECONDS	3600	Page cache TTL
PACKS_DIR	.packs/	Where packs are stored
PORT	3225	HTTP port

Register in any MCP client:

{
  "mcpServers": {
    "source-pack": {
      "command": "node",
      "args": ["/path/to/source-pack-mcp/dist/src/http.js"],
      "env": { "PORT": "3225" }
    }
  }
}

Development

npm run check   # typecheck + lint + build + test
npm test        # run tests

License

MIT