Mykhailo Kushnir
TL;DR
We spent the last few weeks tracking down every government body that has shipped AI into public services and built an interactive index you can browse on a world map.
The headline finding: across roughly fifty serious candidates, only three government bodies on Earth have actually shipped an MCP server or an equivalent agent-addressable endpoint into public preview or production.
- 1France - data.gouv.fr (Feb 25, 2026)
- 2US Government Publishing Office (Jan 22, 2026)
- 3US Census Bureau (Mar 2026)
A further handful (US Treasury, CMS, the French DGFiP tax system) have been announced or built by volunteers against official endpoints, but are not yet officially shipped by the agency itself.
The second story, equally useful, is what governments are shipping for citizens directly: chatbots, assistants, voice agents. Dozens of cases, wildly varying quality. The tracker maps the defensible ones.
Browse the full tracker - with the interactive map
Why we built this
On February 25, 2026, France's open-data agency shipped something quietly historic. They published an MCP (Model Context Protocol) server at mcp.data.gouv.fr/mcp that lets Claude, ChatGPT, and Gemini query French public datasets through natural conversation, directly, without integrations.
The server crashed in its first days. Too much traffic. The French redirect page read sorry.data.gouv.fr.
That moment went viral for a reason. A government had just proven something every SMB founder now has to reckon with: your data, your workflows, your tools are either reachable by AI agents or they are not. The middle ground is shrinking.
Every press office in the world is writing AI strategies. Three have plugged into the wire that ChatGPT and Claude actually speak. We thought that was worth naming, plainly.
How the tracker is organised
Three tiers, in descending order of technical advancement:
- Tier 1: Agent-addressable APIs. An AI model can discover and call these with no custom glue code.
- Tier 2: Citizen-facing AI agents. Chatbots, assistants, voice systems that serve the public directly.
- Tier 3: Back-office AI improving public services. Fraud detection, benefits triage, audit targeting.
Every entry has at least one verifiable primary source. Low-confidence or single-sourced claims are flagged as commonly misreported.
What the data tells you
A few patterns we noticed while compiling it:
Early deployments fail publicly, and the sector still ships the next one. NYC's MyCity told businesses to break the law. It got overhauled and other cities kept shipping. The SSA chatbot has documented glitches. The DWP's fraud AI has documented bias. None of these agencies stopped. The pattern is not "wait until it's safe." The pattern is ship, measure, fix.
The working deployments are narrow. Kela answers benefits questions. Denver's Sunny answers 311 questions. ARCA's Tina recovers fiscal keys. None of these try to be everything. The governments that shipped well picked a bounded domain, locked the scope, and let humans handle the rest.
MCP is the wire, and the wire is here. A year ago "connecting AI to your data" meant hiring an engineer. Today it means installing a server. If you are not thinking about which of your internal workflows should be reachable this way, your competitors either are or will be within the quarter.
What we left out, and why
These show up in coverage and sound credible until you pull the source. We excluded them. Naming what is not real is part of making the list useful.
- Finland - Aurora AI: Launched 2020, ran until end of 2022. Produced a reusable component and a service-network vision, not a live citizen-facing agent.
- Israel - GovAI: Strategy and funding call for 13 pilots. No shipped production chatbot we could verify.
- Spain - ALIA as a citizen chatbot: ALIA is a sovereign LLM that administrations can use to build chatbots. No verified citizen-facing deployment under the ALIA brand yet.
- Germany - AuthorityGPT: OECD describes it as prototype, not deployed product.
- French DGFiP MCP as "government-shipped": The running version is community-built against official endpoints, not sponsored by DGFiP.
- US Treasury and CMS MCP as "live": Listed in AI use-case inventories. No public endpoint yet.
The full list of misreports, with sources, is in the tracker.
Use the tracker
Browse the AI Agent Government Tracker - interactive map, three tier tables, methodology, and the "commonly misreported" sidebar.
If you spot a missing entry, a broken source, or a deployment we got wrong, email mykhailo.kushnir@destilabs.com. The tracker is version 1.0 and we plan to keep it current.
Who we are
DestiLabs builds AI-agent-enabled products and internal tools. We ship for clients the way the governments above ship for citizens: narrow scope, measurable outcome, real infrastructure.
If you read the tracker and thought "I should figure out which of my workflows is MCP-shaped," that is the conversation we have with clients every week. Book a scoping call.