GSA Is Running a Government-Wide Hackathon to Build MCP Servers for Federal Data

The General Services Administration is putting Model Context Protocol to work on one of the messiest problems in government tech: getting AI agents to actually read and write to federal systems. The GSA AI Community of Practice is hosting a virtual, government-wide hackathon running September through November 2026, and it's specifically about building MCP server prototypes for open data assets and service delivery use cases. Industry partners Databricks, OpenAI, and IBM are pitching in on training and mentorship.

If you haven't touched MCP yet, the short version: it's an open-source standard, originally from Anthropic, for connecting AI applications to external tools and data sources. Think of it as a common plug shape so any LLM-based agent can call any tool without a bespoke integration every time. GSA's framing is blunt about why this matters for government specifically: agencies have sunk enormous investment into open data and digital services, but most of it is invisible to AI systems. There's no "structured, context-rich translation layer" that lets an agent query a dataset or submit a form the way a human would through a portal. MCP servers are meant to be that layer.

Registration for the hackathon has already closed, but the structure is worth understanding because it's a template other organizations could reasonably borrow. Participants, who must have a .gov or .mil email address (federal, state, local, territorial, or tribal), pick one of three project types:

• Dataset Access Server – lets an AI app query an open government dataset.
• Service "Read" Integration – lets an AI app retrieve information from a government service.
• Service "Write" Integration – lets an AI app submit information to a government service, sandboxed for obvious reasons.

Teams can go up to five people, mixing engineers with domain experts, or fly solo if you're a software engineer with domain access. GSA is explicit that this isn't engineers-only: understanding the underlying business process, where authoritative data actually lives, and what its quality issues are, counts as much as code.

What Judges Are Actually Looking For

This is the part practitioners should pay attention to, because the rubric doubles as a decent checklist for any MCP server, hackathon or not.

Technical quality is scored across three criteria. First, code readability: logical file structure, self-documenting naming, and separation of concerns between data models, tool definitions, and API interaction. Second, performance and reliability, where submitters have to bring their own evaluation setup, meaning documented test inputs and results, not just a demo that worked once. Judges are told to reward graceful error handling and stable performance across varied inputs, not just a happy-path run.

Third, and probably the most useful guidance if you're new to MCP: adherence to design principles. GSA wants servers built for agents, not humans. That means using Pydantic models (a Python library for validating structured data) for tool inputs, writing tool names and descriptions precise enough that an LLM can pick the right one without ambiguity, and designing tools around actual user goals rather than mirroring REST endpoints one-to-one. In other words, don't just wrap GET /users/{id} and call it a tool. Judges are explicitly told to check whether each tool represents "a real agent-facing action" rather than a thin API pass-through.

Beyond technical quality, submissions get scored on innovation (creative agent workflows score higher), mission alignment (does it actually help service delivery or oversight), and presentation, including demo flow and evidence of testing.

Final deliverables are: a documented GitHub repo (GSA provides a starter template), a slide deck covering the problem statement and technical approach, and a separate evaluation document covering testing methodology, performance metrics, security considerations, and lessons learned. That last requirement, a standalone eval doc, is a nice forcing function. It's easy to skip evaluation when you're racing a hackathon clock, and GSA is baking it into the grading instead of treating it as optional polish.

Time commitment estimates are refreshingly specific: 10 hours for API research and tooling setup, 10 hours for initial build, 20 hours for evaluation and improvements. That 20-hour evaluation number is roughly double the build time, which tells you where GSA thinks the real engineering effort should go.

There's a training cadence attached too. MCP 101 kicks off September 23, followed by vendor sessions from Databricks and IBM through early October, a kickoff ceremony October 13, an MCP 201 session October 29, and weekly GSA office hours from mid-October through November 19. Because of federal legal restrictions, there's no cross-agency cash prize, but winners get recognized at an awards ceremony in front of senior officials.

Key Takeaways

The technical bar here is set higher than the typical civic-hacking event, and that's intentional. GSA isn't asking for demos that impress a room; it wants reusable infrastructure. A working MCP server for, say, a Census Bureau dataset or a benefits-eligibility lookup could plausibly get reused by other agencies facing the same integration gap, which is presumably the point of standardizing on GitHub templates and shared evaluation docs in the first place.

For engineers outside government, the prerequisites list doubles as a solid MCP starter kit regardless of employer: Python 3.10+ or Node.js 18+, familiarity with the MCP SDK, and testing through Claude Desktop, the VS Code MCP extension, or the MCP Inspector tool. Understanding JSON-RPC and HTTP/SSE transport basics is assumed, as is general REST API literacy for anyone touching read/write integrations.

The bigger signal is that a major federal agency is treating MCP not as a novelty but as candidate infrastructure for how AI systems will eventually query and act on government data. Whether any of these prototypes graduate from sandbox to production is an open question. But the rubric, especially the insistence on real evaluation artifacts and agent-native tool design, reads like GSA wants something more durable than a hackathon trophy shelf.