Driving AI readiness through Model Context Protocol (MCP) servers

About GSA

The General Services Administration (GSA) is a federal agency responsible for supporting the basic functioning of other federal agencies by providing centralized procurement services, real estate management, technology solutions, and administrative support across the government.

The challenge

Policymakers, academics, and members of the public increasingly use AI systems to find information online. For many searches, the authoritative source of truth is data collected and disseminated by the federal government. However, federal data, in its current form, is often not easily accessible to AI systems. Consequently, AI systems often point users to less authoritative sources, such as third parties that republish government data that is often outdated or lacks context.

On top of full-time agency responsibilities, five U.S. Digital Corps Fellows—one Software Engineering Fellow and four Data Science and Analytics Fellows—spearheaded groundbreaking work to make federal data AI-accessible, unlocking new ways for the public, researchers, and policymakers to access and innovate with federal data.

The Fellows tested two federal datasets:

1. CDC’s PLACES, a widely-used, open source dataset that provides information about chronic disease and other health-related data, and
2. Department of Treasury’s USAspending, the official open data source on federal spending, and a more complex data environment that aggregates spending data from more than 400 data elements across government

The team prompted different AI systems with a set of questions for which the datasets contained specific, numeric answers. The baseline assessment found that off-the-shelf Large Language Models (LLMs) struggled significantly to accurately retrieve and interpret numeric data from the datasets. In testing the two datasets, they found data retrieval accuracy to be 0% (USAspending) and 2.1% (CDC PLACES).

Neither prompt engineering (which improved ChatGPT’s accuracy to just 4.2%) nor attempts to have the models construct API queries directly (improving accuracy to 17%) meaningfully increased performance to a dependable level, underscoring the need for a more robust approach to reliable data extraction. Without stronger AI retrieval, users making the wide range of decisions informed by federal data risk relying on incomplete, outdated, or non-authoritative sources.

The approach

Currently, the most common way to programmatically access data is through an Application Programming Interface (API). APIs define a structured way to send requests and receive responses between systems. However, due to how much API formats and features vary, LLMs are not capable of directly leveraging them for useful data without assistance. One solution that bridges this gap is the Model Context Protocol (MCP).

The MCP creates a standard for how AI systems can securely and reliably communicate with external resources, including APIs. The Fellows explored this possibility by developing MCP servers for CDC PLACES and USAspending using the Python FastMCP package. By connecting these servers to an AI model through an MCP client, when the model is asked a question, it can invoke the appropriate MCP server to generate an API query and retrieve data from the relevant dataset.

The impact

As AI systems are increasingly relied on for information and data retrieval, the federal government must update its public datasets to match this demand.

Implementing pilot MCP servers improved LLM data retrieval accuracy from a baseline of 0-5% to 95% across both CDC PLACES and USAspending datasets, with response times dropping from minutes to seconds. The Fellows’ proof of concept demonstrates that MCPs offer a practical, scalable path for making authoritative federal data reliably accessible to AI systems, reducing the risk of public reliance on outdated or unofficial sources and providing agencies with a replicable model for AI-ready data modernization across government.

To promote awareness of the potential of MCP technology for AI model data access, the Fellows met with stakeholders across executive agencies, from Chief Data Officers to the Office of the Federal Chief Information Officer at OMB, and presented workshops to working groups including GSA’s AI Community of Practice.

They drafted and published their key findings and recommendations in a report to further promote the potential of this technology across the federal government. The report lays out a strategy to better ensure government data is accessible to the millions of people leveraging LLMs to understand the world around them.