The government's own records — 490+ datasets from 40+ federal agencies — cross-linked by entity, geography, and time since 2010, vintages reconciled, every field normalized, one query away. Ask Claude to run the numbers and show its work, or run the statistics yourself from Python or SQL. Every answer traces back to a primary source.
pip install 'askamerica[engine]'
"Lies, damned lies, and statistics." The fix for the third one is showing the work — every answer here ships with its SQL and its primary source.
Crime, healthcare costs, inflation — the big questions come with competing explanations, each backed by somebody's cherry-picked chart, and the data alone won't crown a winner. What it can do is pin down the checkable part — what happened, where, when, to whom — so the argument moves to the part that's genuinely hard. Framed well, a surprising number of questions have answers sitting in the record. Claude runs the numbers while you watch.
Provenance, reproducibility, entities and geographies that actually join — the properties professionals bill hours against, wherever the work has to be data-driven. One key, pay only for bytes scanned, first GB every month free.
Updated daily from primary government sources, with history back to 2010. All schemas cross-linked by legal entity, geography, and time. No API keys per dataset — one AskAmerica key accesses everything. Browse the full schema, table & column catalog →
Point a well-prompted model at a source it can reach, and it will pull real numbers and run real statistics — that part isn't in dispute anymore. What it won't do on its own: get past a bulk file with no API, a login wall, or a CAPTCHA gate that blocks the data outright; resolve the same organization or person across an SEC filing, an FEC contribution, and a USPTO patent without you doing that identity work by hand; or run an actual similarity search over filing text or patent claims — it'll try, but there's no index for it to search, scoped to one source or five. None of that is a prompting problem — it's a ceiling, and it's there no matter how good the researcher asking is.
From an internal side-by-side evaluation, 2026-08-23 — four personas on one question, scored by an independent judge, who ranked this run's AskAmerica answer first: "the most rigorous single answer... all converging on the same null."
Everyone walking into that meeting already ran the obvious prompt the night before — same summary, same three bullet points. Be the person who brings the number nobody else in the room has, with a query behind it that holds up when someone asks where it came from.
None of this is the model improvising. It's a fixed set of tools, purpose-built for the data, that the model calls the way it calls anything else — so the hard part happened at build time, not at answer time.
Could you build this yourself? Sure — it's software, and given enough time and headcount, anything is buildable. The harder question: are you a software engineer? Are you also a data analyst who's spent twenty years learning exactly where government data lies to you? Most people are one or the other. This took both, for years, before the first query ever ran.
AskAmerica handles ingestion, formatting, and serving. You get a key, connect a client, and query — three steps, no data to deploy or maintain.
AskAmerica is in preview while we finish filling out data coverage. Every dataset is included, on the free tier, for everyone — the only thing we ask in return is your feedback. Paid plans below open up once preview ends.
Same data, same API key — choose the client that fits your workflow.
Bundled JRE — no Java, no Python, no config files. Open the app, enter your API key, click Configure. Works with Claude Desktop and any other MCP-compatible client.
Or via pip: pip install 'askamerica[mcp]' then askamerica mcp-config.
list_schemas → what datasets are available? list_tables → what tables are in fec? describe_table → what columns does contributions have? query → run any SQL and return results
Claude calls these automatically — you just ask questions in plain English. No SQL required on your end.
One package, zero boilerplate. Returns a pandas DataFrame or a raw JDBC connection — JPype is managed internally, never exposed.
pip install 'askamerica[engine]' askamerica login # set your API key
Python 3.8+. The engine JAR downloads automatically on first query (cached under ~/.askamerica). On Python 3.10+ a JVM is bundled too; on 3.8–3.9, Java 11+ must be on your PATH.
import askamerica as aa df = aa.query(""" SELECT company_name, value_dollars FROM sec.financial_facts WHERE canonical_name = 'Revenue' ORDER BY value_dollars DESC FETCH FIRST 10 ROWS ONLY """) print(df)
conn = aa.connect()
rs = conn.createStatement().executeQuery(
"SELECT cik, company_name FROM sec.filing_metadata"
" ORDER BY filing_date DESC FETCH FIRST 5 ROWS ONLY")
while rs.next():
print(rs.getString("company_name"))
conn.close()
A single fat JAR — use from Node, Go, Java, DBeaver, or any JDBC client. No other dependencies.
↓ askamerica-engine.jarJDBC URL: jdbc:askamerica:source=geo,sec Driver JAR: askamerica-engine.jar Driver class: org.apache.calcite.adapter .askamerica.AskAmericaDriver Username: (leave blank)
const { createConnection } = require('jdbc'); const conn = await createConnection({ url: 'jdbc:askamerica:source=sec,geo', drivername: 'org.apache.calcite.adapter .askamerica.AskAmericaDriver', classpath: ['./askamerica-engine.jar'] }); const rows = await conn.query( 'SELECT cik, company_name FROM sec.filing_metadata FETCH FIRST 5 ROWS ONLY');
// go-jdbc wraps the JVM via CGo db, _ := sql.Open("jdbc", "jdbc:askamerica:source=sec,geo") rows, _ := db.Query( "SELECT cik, company_name FROM sec.filing_metadata FETCH FIRST 5 ROWS ONLY")
AskAmericaDriver driver = new AskAmericaDriver(); Connection conn = driver.connect( "jdbc:askamerica:source=sec,geo,econ", new Properties()); ResultSet rs = conn.createStatement().executeQuery( "SELECT cik, company_name FROM sec.filing_metadata FETCH FIRST 5 ROWS ONLY");
A politician says renewable energy is driving up your utility bill. A headline says a new policy caused a crime wave. Instead of taking it on faith, click the AskAmerica icon in your toolbar. It pulls apart every checkable claim on the page and tests each one against government data — census, energy, crime, health, hundreds more datasets — then hands back a verdict: true, false, partially true, or not checkable, with the actual numbers and sources behind each call. No more "sounds right" — an actual answer, in about the time it takes to read the headline.
Chrome/Edge/Brave: unzip, open chrome://extensions, enable Developer mode, click Load unpacked, select the unzipped folder.
1. You're on a page with a claim worth checking 2. Click the AskAmerica toolbar icon 3. "Validate with AskAmerica" opens your own Claude session — no API call from the extension itself 4. It tests every claim against real government data and publishes a report: verdict, evidence, sources 5. Back on the page, each claim is highlighted by verdict — green true, red false, amber partial — with the full breakdown one click away
Runs on your existing Claude Desktop or claude.ai plan — not a separate API key, and the extension itself never talks to a model. Requires the AskAmerica MCP engine running locally (see the MCP tab) to read and publish reports.
Enter your email. We'll send a one-time code — no password required.
Your free API key is below. Keep it safe — it grants access to your quota.
pip install 'askamerica[engine]'askamerica login # key: import askamerica as aadf = aa.query("SELECT 1")