corroborate-mcp

An MCP server that tells AI agents how independently corroborated a claim is.

Give it a claim, get back an honest verdict: how many independent story origins are reporting it — not how many articles mention it. An AP story echoed by 40 outlets counts as one origin. That judgment layer is what raw search doesn't give you.

"NASA delayed the Artemis III landing"  →  CONFIRMED    (4 independent origins, confidence 0.9)
"Zorblatt Corp acquired Portugal"       →  UNCORROBORATED (0 origins, confidence 0.05)

What it measures — and what it doesn't: reporting corroboration, not truth. A claim every outlet syndicated from one wire story comes back SINGLE_SOURCE. A claim nobody covers comes back UNCORROBORATED even if true. The verdict says exactly what was measured, flags its own weaknesses (notes), and discloses when a source engine was down (coverage: "degraded") so absence of evidence is never silently confused with evidence of absence.

• Keyless & free to run locally — Google News RSS + GDELT + Hacker News. No API keys, no accounts.
• Deterministic — no LLM calls. Same evidence in, same verdict out. Typical call: 1–3s (10s worst case when an engine lags; repeat queries are cached ~10 min).
• Agent-first output — small, tight JSON an agent can act on directly.

Quickstart

Requires Node ≥ 18.

Claude Code

claude mcp add corroborate -- npx -y corroborate-mcp

Claude Desktop / Cursor / any MCP client (claude_desktop_config.json, .cursor/mcp.json, …)

{
  "mcpServers": {
    "corroborate": { "command": "npx", "args": ["-y", "corroborate-mcp"] }
  }
}

Straight from a checkout

npm install
node cli.js "The Federal Reserve held interest rates steady"   # engine smoke test, no MCP client needed

Tools

corroborate_claim

The main event. { claim, window_days?=7, max_sources?=8 } →

{
  "claim": "Acme Corp acquires Widget Industries for $2 billion",
  "corroboration": "CONFIRMED",
  "n_independent_sources": 3,
  "confidence": 0.9,
  "coverage": "full",
  "sources": [
    {
      "headline": "Acme Corp to acquire Widget Industries in $2 billion deal",
      "outlet": "Financial Times", "domain": "ft.com", "url": "https://…",
      "published_at": "2026-07-09T14:02:00Z", "age_days": 0.9,
      "wire": false, "echoed_by_n_domains": 1
    }
  ],
  "notes": ["syndication detected: near-identical headlines across multiple domains were collapsed into one origin"],
  "method": "reporting-corroboration: counts INDEPENDENT story origins (syndication-aware), not raw article count; does not adjudicate truth"
}

Field	Meaning
corroboration	CONFIRMED (≥2 independent origins) · SINGLE_SOURCE (1) · UNCORROBORATED (0)
n_independent_sources	Distinct story origins after syndication clustering — not article count
confidence	0–1. Scales with origins and cross-engine agreement; penalized for wide single-origin echo
coverage	full, or degraded when a source engine failed this call — a weak verdict then may reflect the outage, not the claim
sources[].wire	Origin is a wire service (AP/Reuters/AFP/UPI)
sources[].echoed_by_n_domains	How many domains carried this same story
notes	Honest caveats: syndication collapse, breaking-news cascade (<24h coverage), relevance-gate discards, engine outages

find_sources

Cheaper primitive: { query, window_days?=7, max_sources?=10 } → deduped multi-engine article list (outlet, domain, date, engine) with no verdict. Use it when you want the evidence but your own judgment.

How the independence judgment works

1. Search Google News RSS, GDELT, and Hacker News in parallel (keyword extraction from the claim; quoted phrases preserved).
2. Relevance gate — an article only counts if its headline genuinely addresses the claim (≥2 core-token hits AND ≥35% of core tokens). Loose topical echoes are discarded and the discard count is reported.
3. Syndication clustering — near-identical headlines across different domains (Jaccard similarity ≥ 0.55) collapse into one origin; wire domains are flagged.
4. Verdict — independent origins = distinct clusters. Confidence rises with origins and cross-engine agreement, falls for wide single-origin echoes. Every weakness the engine knows about goes in notes.

Measured accuracy — published because you should demand it from any fact-check tool

Run against a 40-claim labeled benchmark (benchmark/ — claims, labels, category definitions, and harness are public; re-run it yourself with npm run benchmark). Latest run 2026-07-12:

Metric	Result
Fabricated claims falsely CONFIRMED	0/10 (fabricated = entity has no real referent)
Widely-reported true claims missed	0/12 (12/12 CONFIRMED)
Niche true claims detected	7/8
Stale claims correctly windowed	3/5
Distorted claims falsely CONFIRMED	4/6 — read the warning below
Latency	p50 0.4s · p95 ≤8s (engine-outage worst case)

⚠️ The distorted-claim number is the one to respect. This tool measures whether independent reporting exists around a claim's topic and entities — it does not do stance detection. A distorted version of a real event ("OpenAI released GPT-6" when the real news is GPT-5.6; "the EPA strengthened rules" when it weakened them; "SpaceX's Starship exploded on the pad" when the real coverage is a routine launch) can come back CONFIRMED because real coverage token-matches it. If your input may be adversarial or detail-critical, treat CONFIRMED as "this topic has independent coverage — now verify the specific details against the returned sources." Stance detection is the v1.1 roadmap item.

Also measured: claims about recurring events (championships, elections) can match the current cycle's coverage — "Argentina won the World Cup in Qatar" (2022, true) confirms against 2026 tournament coverage — and old events with fresh retrospective/anniversary coverage can return SINGLE_SOURCE.

Honest limitations

• No stance detection — see the measured 4/6 above. CONFIRMED means independently covered, not independently verified in every detail.
• English-language, headline-level analysis. Paywalled body text is not fetched.
• Recency-biased: the default window is 7 days (max 90). Old claims come back UNCORROBORATED — that's a window statement, not a falsity verdict.
• Two outlets independently rewriting the same wire story can occasionally slip past clustering; genuinely different phrasings of one origin may occasionally count as two.
• If all three engines are down, tools return an error — never a fake UNCORROBORATED.

Development

npm test              # golden-claim suite — deterministic, no network
npm run test:mcp      # MCP stdio handshake + tools/list
npm run test:live     # live invariants against real engines
npm run benchmark     # 40-claim labeled accuracy benchmark (live, ~2 min)
npm run test:release  # all of the above — required green before every release

MIT © Ezra Cohen