chefcohen/corroborate-mcp

**The honest corroboration layer for AI agents.** Hand it a claim; it tells you how *independently* that claim is actually being reported — and shows its work.
> An AP story echoed by 40 outlets is **one** origin, not 40 citations. Other verifiers count citations; Corroborate counts independent *origins* — syndication-aware — then publishes its own error rates and discloses coverage gaps. That's the judgment layer raw search and citation-checkers skip.
When an agent should call this
- Before **relying on a current-event / news claim** — is it independently reported, or one source echoed everywhere?
- To **rank or filter sources** by genuine independence, with a confidence score.
- When you need a **verdict you can defend** — every answer ships its evidence, its confidence, and its own caveats.
Not a web search engine (use Exa/Tavily/Brave for that) and not a human fact-checking network. It's the narrow, honest primitive in between: *how independently is this reported?*
Why it's trustworthy — by design
| | |
|---|---|
| 🔑 **Zero API keys, zero accounts** | Keyless engines (Google News RSS · GDELT · Hacker News). Nothing to sign up for. |
| 👁 **Read-only** | It reads public news. It cannot act, write, spend, or touch your data. |
| 🎯 **Deterministic** | No LLM in the loop — same evidence in, same verdict out. p50 ~0.4s. |
| 📖 **Open source (MIT)** | Read every line of the judgment. |
| 📊 **Published error rates** | We measure our own false-CONFIRMED / false-UNCORROBORATED rates on a public benchmark and print them below. Most "fact-check" tools don't. |
| 🚫 **Never lies by omission** | If sources are unreachable it returns an **error**, never a fake "UNCORROBORATED." |
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 states exactly what was measured, flags its weaknesses in notes, and sets coverage:"degraded" when an engine was down — so absence of evidence is never quietly sold as evidence of absence.
Quickstart
Requires Node ≥ 18. No keys, no config.
**Claude Code**
**Claude Desktop / Cursor / any MCP client** (claude_desktop_config.json, .cursor/mcp.json, …)
**Try it with no MCP client**
Tools
corroborate_claim — the verdict
{ claim, window_days?=7, max_sources?=8 } →
| 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. Rises with origins + cross-engine agreement; penalized for wide single-origin echo |
| coverage | full, or degraded when an engine failed this call — a weak verdict may then 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), relevance-gate discards, engine outages |
find_sources — the raw evidence
{ query, window_days?=7, max_sources?=10 } → deduped multi-engine list (outlet, domain, url, date), no verdict. The cheaper primitive when you want the coverage and your own judgment.