citation-intelligence

Citation Intelligence MCP

A free, self-hosted MCP server that tells your agent what LLMs cite - across Perplexity, Google AI Overviews, ChatGPT, Claude, Gemini, and Bing.

What this is

An MCP server for agents and developers who need to know which URLs get cited by AI search engines for any query. Install once, query from any MCP-compatible client (Claude Desktop, Cursor, Claude Code, Continue, Cline, n8n, LangGraph). Self-hosted, no account, no centralized backend. Bring your own API keys; nothing is stored on a remote server.

Who this is for

Install this if you're:

• Building an agent that does research and want it to cite sources LLMs already trust
• A solo dev or indie hacker checking whether your SaaS is showing up in AI search
• A content creator confirming your articles are being cited by ChatGPT, Claude, or Perplexity
• An SEO or GEO practitioner who wants programmatic citation data without a $295-$499/mo dashboard
• Running an editorial pipeline and want citation-deficit-driven topic selection
• Comparing competitor visibility across AI engines for any niche

Do NOT install this if you want:

• A polished marketing dashboard with charts and team seats - try Profound, AthenaHQ, or Otterly.AI
• A hosted service with SLAs - this is self-hosted by design
• Citation tracking for academic papers - try citecheck
• 350M+ pre-modeled prompts - that's Ahrefs Brand Radar

Why this exists

The AI citation tracking market is dominated by VC-funded dashboards starting at $295/mo. None ships MCP-first. If you're an agent or developer who wants citation data piped directly into your workflow - not into a SaaS login - there isn't a tool for you. This is that tool.

Tools

Tools are grouped into seven namespaces: citations_*, domain_*, signals_*, panel_*, report_*, competitors_*, audit_*. The prefix is the question category; the suffix is the action. Wire names use underscores (not dots) so Anthropic-API-based MCP clients (Claude Desktop, Claude Code) can forward the tool list without HTTP 400.

Start with citations_provenance or domain_am_i_cited. Single-engine results (citations_check with a pinned engine) are directional; multi-engine consensus is the honest signal. A URL cited by 4 of 5 engines is a very different finding than one cited by 1.

citations_* — query-level: who cites what, with what evidence

domain_* — domain-level: am I cited, what for

signals_* — external signals: AI Overview, Wikipedia, GSC, answer-box position

panel_* — saved query panels (editorial watchlists)

report_* — turnkey reporting artifacts

competitors_* — competitive landscape per query

audit_* — fixable on-page / on-site checks

Prompts

Server-side prompt templates the client can offer end users (call via the MCP prompt list):

• audit_citation_readiness(url) - chains citations_predict + audit_schema
• audit_competitor_snapshot(query, your_url?) - chains competitors_canonical_set + competitors_compete
• audit_crawler_checkup(url) - runs audit_crawler_access and writes a remediation list
• audit_gap_analysis(domain, days?) - drives signals_gsc_gap and suggests next moves
• audit_sitemap_coverage(sitemap_url) - runs audit_sitemap_map and recommends priorities

Resources

Cache views the client can read or subscribe to (no tool call required):

• citation://cache/summary - entry counts by type/engine, unique queries/URLs, oldest/newest
• citation://panels - saved panels + per-panel snapshot counts
• citation://docs/llms-txt - llms.txt primer (markdown)
• citation://docs/ai-crawlers - AI crawlers cheatsheet (markdown)
• citation://domain/{domain}/cited-for - dynamic template: citations for {domain}

What this actually measures

Every response includes a surface field that tells you exactly how the data was collected. Understanding this is important before drawing conclusions.

Per-engine notes

perplexity (consumer_scrape) — Sonar Pro via the Perplexity API with a consumer-equivalent system prompt. Reasonably close to Perplexity.ai. Citations come from search_results in the response; the citations fallback contains URL-only entries without title.

claude (api_proxy) — Claude Sonnet via the Anthropic Messages API with web_search tool enabled. The consumer Claude.ai product uses different routing and ranking logic. Citation behavior can differ, especially for recent/time-sensitive queries.

openai (api_proxy) — gpt-4o + the web_search_preview tool via the OpenAI Responses API. Replaces the deprecated gpt-4o-search-preview alias OpenAI retired; base gpt-4o plus the tool is the supported path.

gemini (api_proxy) — Gemini 2.5 Pro via the Generative Language API with google_search grounding. Consumer Gemini uses the same grounding index but different re-ranking. Results are directional.

google_ai_mode (consumer_scrape) — Google AI Mode results via SerpAPI. Closest to what users see in Google Search. Requires SERPAPI_KEY.

bing_serp / brave_serp (web_rank) — Traditional SERP rank. Does NOT measure LLM citations. Use citations_check with these engines to compare organic web rank against LLM citation rank. domain_am_i_cited refuses these engines — it only measures LLM behavior.

The proxy nature of api_proxy engines is a feature, not a bug: it lets you run citation checks without consuming expensive consumer-product quota. Just don't report API-proxy numbers as "ChatGPT cites you" without the caveat.

Every tool response includes an interpretation_note field that summarizes the fidelity in one sentence. Full per-engine fidelity ratings: docs/surface-fidelity.md.

Quick start

npx -y @automatelab/citation-intelligence

Requires Node 20 or later.

Claude Desktop

Add to %APPDATA%\Claude\claude_desktop_config.json (Windows) or ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "citation-intelligence": {
      "command": "npx",
      "args": ["-y", "@automatelab/citation-intelligence"],
      "env": {
        "PERPLEXITY_API_KEY": "pplx-...",
        "SERPAPI_KEY": "...",
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "OPENAI_API_KEY": "sk-...",
        "GEMINI_API_KEY": "..."
      }
    }
  }
}

Set only the keys you have. Any MCP client that supports stdio transport works - same command / args pattern.

How it stays free

• No central backend. The server runs on your machine. Nothing is uploaded.
• Free tier first. SerpAPI gives 100 free Google AI Overview lookups/month. Bing Web Search has a free tier. Perplexity offers free Sonar access on signup.
• Bring your own paid keys if you want the premium engines (Claude, ChatGPT, Gemini). Keys pass through to the vendor and never touch any third party.
• Local cache at ~/.config/citation-intelligence/cache.json. Repeated queries hit cache, not API. Default TTL: 7 days.
• citations_predict runs with zero keys - it scores citation likelihood from public signals (Wikipedia, schema.org, llms.txt, GitHub) without firing any LLM.

Privacy

• All API calls go from your machine directly to the vendor (Anthropic, OpenAI, Google, Perplexity, Bing, SerpAPI).
• No proxy. No analytics. No telemetry by default.
• API keys are read from environment variables on the MCP process - never logged, never persisted.
• Cache file lives at ~/.config/citation-intelligence/cache.json. Delete it any time.

Environment variables