Economic statistics your AI can actually cite.

Why citations, why now

AI agents write a growing share of the world's reports, briefs and analysis, and they miscite numbers at measurable rates. StatCite exists so the number in your draft is the number in the official series, with the receipt attached.

We measured it ourselves, on our own claim

The stats above are general research on AI and citations. This is StatCite's own benchmark: 6 models across 3 vendors (4 Anthropic models, 1 OpenAI, 1 Google preview model) answering 100 real economic-statistics questions, instructed to answer from memory only. No tools, no retrieval, no StatCite. It's the honest baseline StatCite exists to fix.

*claude-haiku-4-5's Answer Rate falls below the 70% threshold the pre-registered methodology requires for a model's result to be interpreted at all. It is reported for completeness, not as a headline finding, per the methodology's own pre-committed rule.

Within-Tolerance Rate, Confabulation Rate, and Answer Rate are a single unit by our own publication covenant, never quoted apart from one another. At this sample size (Wilson 95% CIs, minimum detectable model-to-model difference ~12–15pp) this is not a league table. How the benchmark is built, why it is hard to game, and what is pre-registered for the next run: the benchmark page. Full methodology, pre-registration, raw model outputs, and scoring code: the R1 report. Post-publication sensitivity analyses and the logged protocol deviations for this run: the R1 addenda.

Connect your agent

One URL. No signup, no API key, no OAuth. Works in every major MCP client.

Once connected, three ready-made prompts appear in clients that support them: fact_check (paste a draft, every macro claim in it gets checked), country_brief and cite_this_stat. What they do.

Name:  StatCite
URL:   https://statcite.com/mcp

MCP server URL:  https://statcite.com/mcp
Authentication:  No authentication

{
  "mcpServers": {
    "statcite": { "url": "https://statcite.com/mcp" }
  }
}

{
  "servers": {
    "statcite": { "type": "http", "url": "https://statcite.com/mcp" }
  }
}

/plugin marketplace add asokore/statcite
/plugin install statcite@statcite

claude mcp add --transport http statcite https://statcite.com/mcp

gemini extensions install https://github.com/asokore/statcite

{
  "mcpServers": {
    "statcite": {
      "command": "npx",
      "args": ["-y", "mcp-remote@latest", "https://statcite.com/mcp"]
    }
  }
}

curl "https://statcite.com/v1/indicator/inflation_cpi?country=BRB&latest_only=true"
curl "https://statcite.com/v1/verify?indicator=inflation_cpi&country=USA&period=2023&value=4.1"
curl "https://statcite.com/v1/fx?amount=100&from=USD&to=BBD"

What a verified number looks like

An agent drafts “US inflation was 4.1% in 2023.” Before publishing, it calls verify_stat:

// tools/call → verify_stat {indicator:"inflation_cpi", country:"USA", period:"2023", claimed_value:4.1}
{
  "verdict": "match",
  "claimed_value": 4.1,
  "official_value": 4.116338383,
  "explanation": "Claimed 4.1 vs official 4.1163 % (annual average) for 2023, consistent (difference of 0.016 pp is within normal rounding).",
  "citation": {
    "source": "World Bank",
    "dataset": "World Development Indicators",
    "series_id": "FP.CPI.TOTL.ZG",
    "series_name": "Inflation, consumer prices (annual %)",
    "source_url": "https://data.worldbank.org/indicator/FP.CPI.TOTL.ZG?locations=USA",
    "license": "CC BY 4.0",
    "retrieved_at": "2026-07-25",
    "citation_text": "World Bank, World Development Indicators, series FP.CPI.TOTL.ZG (Inflation, consumer prices, annual %), data last updated 2026-07-13. Retrieved 2026-07-25 via StatCite (https://statcite.com). https://data.worldbank.org/indicator/FP.CPI.TOTL.ZG"
  }
}

Illustrative response. Live values, source-update dates and retrieval dates change as sources publish revisions.

And when the number is wrong, the diagnostics say how it went wrong. Claim 0.041 instead and:

// tools/call → verify_stat {indicator:"inflation_cpi", country:"USA", period:"2023", claimed_value:0.041}
{
  "verdict": "mismatch",
  "claimed_value": 0.041,
  "official_value": 4.116338383,
  "diagnostics": ["The claimed value is ~100× smaller than the official figure, possibly a percent-vs-decimal mix-up (e.g. 0.05 vs 5%)."]
}

Wrong year? Percent-vs-decimal slip? Billions-vs-millions? Sign flipped on a fiscal balance? verify_stat diagnoses the classic errors and returns the correct, citable figure.

Try a live verification

No setup, no key. This form calls the same public GET /v1/verify endpoint your agent would. Pick an indicator, claim a number, get the verdict and the citation.

Twelve tools, one contract: every number carries its receipt

Available over MCP and mirrored as a REST API. All read-only.

Sources & licenses

StatCite is a citation layer over official statistics, never a black box. Data is fetched live from the sources and cached briefly for politeness.

The complete licence ledger, with every source, its verbatim licence basis, verification date and required attribution, is served live at /sources and GET /v1/sources.

Pricing

Free while the agent economy grows up. Infrastructure costs are near zero by design. When paid tiers arrive, the free tier stays.

Support StatCite

StatCite is free, with no ads, no trackers and no data resale. Running it costs very little by design, so there is no funding drive here. If it caught a wrong number before you published it, you can put something in the tin.

Contributions go to the domain, the data-source costs and the time to keep the registry curated. They buy no priority support, no influence over which sources StatCite trusts, and no change to how verify_stat scores a claim. The verdict is not for sale. That is the whole point of the project.

FAQ