evidoria

← Back to browse

Good practice Imported

MCP for Federal Open Data — U.S. Digital Corps Pilot Lifts AI Query Accuracy from Near-Zero to 95%

United States of America · Washington, D.C. · See the United States of America profile · See the Washington, D.C. profile

Evidence: Observational / pre–post Top 7% 87/100 · Ask Evidence Copilot about this practice

A federal fellows pilot found ChatGPT, Gemini and Claude answered USASpending and CDC PLACES questions with near-0% and 2.1% accuracy unaided, jumping to 95% once agencies exposed the same open datasets through Model Context Protocol servers.

0 % (near-zero)
USASpending query accuracy without MCP
2.1 %
CDC PLACES query accuracy without MCP
95 % (approx., both datasets)
Query accuracy with MCP servers
MCP for Federal Open Data — U.S. Digital Corps Pilot Lifts AI Query Accuracy from Near-Zero to 95%

Details

Maturity
Pilot
Promoter
U.S. Digital Corps (General Services Administration) with the Census Bureau and Government Publishing Office
Period
2025–2026
Keywords
federal statistics, open data, AI infrastructure, spending transparency

Context

U.S. Digital Corps fellows placed inside federal agencies, including one at the Commerce Department's Office of the Under Secretary for Economic Affairs, ran a pilot testing how accurately general-purpose AI assistants (ChatGPT, Gemini and Claude Sonnet 4) could answer questions against two open federal datasets, USASpending.gov and CDC PLACES.

Activities

The models were first queried without any special connector, then re-tested using Model Context Protocol (MCP) servers that exposed the same datasets directly to the AI tools.

Results

Queried unaided, the models answered USASpending questions correctly close to 0% of the time and CDC PLACES questions correctly only 2.1% of the time; with MCP servers exposing the same datasets, accuracy rose to roughly 95% across both datasets, and a documented Worcester County health-statistic query that had previously failed returned the correct figure immediately, complete with confidence intervals and population data.

Conclusions

Following the pilot, the Census Bureau and the Government Publishing Office each published a public MCP server on GitHub, and CMS and the Treasury Department listed MCP deployments in their 2025 agency AI inventories, though officials such as the Census Bureau's chief innovation officer cautioned that MCP access "is not the answer, it's one of many answers" and that coverage remains partial.

Implementation
Indicative cost
Low (< €50k)
Time to results
Short (< 1 year)
Staffing & skills
U.S. Digital Corps fellows embedded in agencies (e.g., the Commerce Department), working with the Census Bureau and the Government Publishing Office.

Conditions for success

  • Publishing MCP servers as open source on GitHub so other agencies could adopt the same pattern.
  • Testing multiple AI models (ChatGPT, Gemini, Claude Sonnet 4) against the same datasets before and after exposing them via MCP.

Common failure modes

  • Agencies themselves describe coverage as partial rather than comprehensive.
  • No dedicated long-term funding line for federal MCP maintenance is documented.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful