GSA Launches Government-Wide MCP Hackathon to Connect AI Agents With Federal Data

The GSA AI Community of Practice is running a government-wide hackathon to build MCP server prototypes for federal open data and service delivery, and it's a useful signal of where agencies think AI infrastructure investment needs to go next. The event runs virtually from September through November 2026, open to any government employee with a .gov or .mil email address (federal, state, local, territorial, or tribal). Registration has already closed, so this is now a matter of watching what comes out the other end.

Model Context Protocol, if you haven't run into it yet, is the open-source standard for wiring AI applications into external systems. Think of it as the connective tissue between an LLM and the actual tools or data it needs to act on, rather than just talk about. It's been picking up steam across the industry as agents move from "answer my question" chatbots toward systems that actually do things: query databases, call APIs, submit forms. The problem GSA is trying to solve is specific to government: there's a massive amount of public data and service infrastructure sitting behind APIs, but none of it has the structured, agent-readable translation layer that would let AI systems query, interpret, or act on it safely.

That's the gap this hackathon targets. Participants pick one of three tracks:

• Dataset Access Server — build a server that lets AI applications query an open government dataset.
• Service "Read" Integration — expose read access to a government service through an MCP server.
• Service "Write" Integration — enable AI to submit information to a government service, sandboxed for obvious reasons.

Teams (solo or up to five people, ideally a mix of engineers and domain experts) get a six-week development window with training, mentorship, and sandbox environments provided by GSA and its industry partners: Databricks, OpenAI, and IBM.

How the judging actually works

This is the part worth paying attention to if you're thinking about entering, or just curious about how GSA wants MCP servers built in practice. Submissions get scored on four axes, and the technical quality criteria are detailed enough to double as a decent internal style guide for anyone building MCP servers at their own org.

• Code readability and maintainability: judges want logical file structure, self-documenting naming, and clean separation between data models, tool definitions, and API interaction logic. Comments should explain intent, not restate what the code already says.
• Server performance and reliability: submitters have to bring their own evaluation setup, documenting test inputs and results. Graceful error handling and stable performance across varied inputs score higher than a demo that only works on the happy path.
• Adherence to MCP design principles: this is the most opinionated criterion, and it's a good checklist for anyone building MCP tooling generally. Servers should use structured input validation (the guidance specifically calls out Pydantic models), tool names and descriptions precise enough that an LLM can pick the right one without ambiguity, and tools designed around actual user goals rather than thin wrappers around individual REST endpoints. In other words: don't just expose your API 1-to-1 and call it an MCP server. Design for what an agent is actually trying to accomplish.

Beyond technical quality, judges weigh innovation (creative use of MCP tooling or novel agent workflows), mission alignment (does it actually improve service delivery or oversight), and presentation quality, including demo flow and evidence of testing.

Final deliverables are a GitHub repo built from a provided template, a presentation deck covering the problem statement and technical approach, and an evaluation document covering testing methodology, performance metrics, security considerations, and lessons learned. GSA has published sample templates for all three, which is a nice touch for people navigating MCP for the first time.

There's no cross-agency monetary prize on the table, due to federal legal restrictions on cash awards across agency lines, but winners get recognized at an awards ceremony with senior government officials in attendance. For a government hackathon, that's about as good as it gets structurally.

The schedule itself is fairly dense for a six-week sprint: an "MCP 101" session kicks things off September 23, followed by two rounds of vendor training from IBM and Databricks in late September and early October, weekly office hours running from mid-October through mid-November staffed by GSA, Databricks, IBM, and Prefect, a kickoff ceremony October 19, a halfway check-in November 3, and an MCP server governance session led by Harness on November 12. Closing ceremony is penciled in for December 16, though GSA notes dates are still subject to change.

On the skills side, GSA is explicit that this isn't engineers-only. Domain experts who understand their agency's data, workflows, and quality issues are treated as equally essential to good server design as the people writing code. That's a sensible framing: an MCP server is only as useful as the mapping between what an agent needs and what the underlying system actually exposes, and that mapping requires someone who actually knows the data. For the technical side, the baseline expectations are pretty standard modern dev environment stuff: comfort with a shell, Git, Python 3.10+ or Node.js 18+, virtual environment and package management (venv, conda, or uv for Python; npm or yarn for Node), and familiarity with MCP-specific tooling like the MCP SDK and a testing client such as Claude Desktop, the VS Code MCP extension, or MCP Inspector.

Key Takeaways

This hackathon is a prototyping exercise, not a production rollout. Deliverables stay in sandbox environments and won't be publicly deployed, and the scope is explicitly limited to data that's already publicly available. But the structure here, cross-agency, multi-vendor, with a genuinely detailed rubric for what a "good" MCP server looks like, is a useful reference point even outside government. If you're building MCP servers anywhere, the judging criteria around tool design (goal-oriented, not endpoint-mirroring) and evaluation documentation are worth stealing regardless of who's grading you. Questions about the program go to mcp@gsa.gov, and GSA says it'll post updates on final dates as they firm up.