The agent layer
An agent is only as good as the map it navigates by.
Software is starting to act on people's behalf in front of their governments. To do that at all, it has to work out which government it is dealing with. That map is being drawn right now, and the question is whether it ends up public or privately owned. We are building the public one.
The questionWho do I tell about a broken streetlight on Linden Avenue in Carpinteria, California?
FIG. 1: THE ANSWER IS NOT THE CITY. IT IS A LIGHTING DISTRICT MOST RESIDENTS HAVE NEVER HEARD OF, AND IT IS A REAL ENTITY IN THE UNGOVR ATLAS. ONE EXAMPLE OF THE SAME RESOLUTION THE ATLAS PERFORMS ACROSS 200+ COUNTRIES.
Why the depth is the point
Everyone has the top of the map. Almost nobody has the bottom.
Any search engine can tell you your city hall. The layer that decides most real questions is the one below it: the water district, the fire district, the school district, the lighting district. In California alone these special districts outnumber the cities by more than ten to one, and most countries have their own version of that layer, whether it is a French syndicat, an English parish council or a Japanese special ward. An assistant that stops at the city gives a confidently wrong answer, which is worse than no answer, because the person stops looking.
The examples on this page are Californian, and they are only examples. The UnGovr Atlas holds 330,000+ government entities across 200+ countries, and the same resolution runs anywhere it reaches. We use California to show the work because it is where we hold enough address data to test ourselves precisely, not because it is where the problem lives.
The agent works for the person
- Anyone can build an assistant that reaches government correctly, not only the companies that can afford a licence.
- The answer can be checked, because records link back to the official sources they came from.
- A town of 800 is in the map on the same terms as Los Angeles, because nothing has to be bought to put it there.
Reaching your government becomes a rental
- Access to the governments you fund and vote for is metered by whoever owns the index of them.
- Coverage follows commercial value, so the places with the least capacity are mapped last or not at all.
- Nobody outside can audit what the map says about a government, or what it quietly leaves out.
And the address will mislead you
The city in the address is not the city with jurisdiction.
Ask a general-purpose assistant who governs an address and it will usually read the city or the postal code straight out of the address and answer with that. A mailing address names the delivery office, not the government that serves you, and the two were drawn by different people for different reasons. Postal geography and administrative geography come apart wherever you look, not only in California: unincorporated areas addressed with the name of the city next door in the United States, a French commune addressed by its bureau distributeur, a UK postcode straddling two council boundaries. The address is a delivery instruction, and it keeps getting mistaken for a jurisdiction.
Montecito is the famous example. Its addresses read “Santa Barbara, CA 93108” and it is not in the City of Santa Barbara at all. It is unincorporated county, with its own fire district, its own water district and its own sanitary district, none of which the city can help you with. Nothing in the address says so.
The testOf the California addresses written “Santa Barbara, CA”, how many are inside the City of Santa Barbara? We hold the addresses and we hold the city boundary, so this is a measurement rather than an opinion.
FIG. 2: 40,277 ADDRESSES WRITTEN “SANTA BARBARA, CA”, TESTED AGAINST THE CITY BOUNDARY, AUGUST 2026. THE OUTSIDE SET IS LED BY THE 93111, 93110 AND 93108 ZIPS. THE SAME TEST PUTS 21.5% OF “CARPINTERIA, CA” ADDRESSES OUTSIDE THAT CITY. ONE PLACE, MEASURED BECAUSE WE HOLD THE ADDRESSES THERE. THE DIVERGENCE ITSELF IS EVERYWHERE.
Two in five is not an edge case, and an assistant parsing the address string gets every one of them wrong while sounding certain. The failure is quiet: the person is told to contact a city that has no authority over their problem, and they believe it. Some of these are not even close. Hollister Ranch Road writes “Santa Barbara, CA” from out on the Gaviota coast, most of the way to the county line.
There is only one way to answer this correctly, and it is not a better language model. It is to put the point on a map and ask which boundaries contain it. That is what the UnGovr Atlas is for, and it is why we draw boundaries rather than just listing agencies: 160,000+ of them so far, in the countries we have reached.
What exists today
Three ways in, at three different depths.
We would rather publish the real state of this than a rounded-up version. The map itself is deep and finished enough to answer the question above. The agent interface on top of it is not there yet, and here is exactly where it stands.
UnGovr Atlas
330,000+ government entities across 200+ countries, 160,000+ of them with boundaries drawn on the ground. Each layer, down to the lighting district in Fig. 1.
How the UnGovr Atlas works →The data API
The same map as structured data, entity by entity, with the parent chain, boundary and available products on each record. This is the interface that currently reaches the full depth.
API reference →The MCP server
An assistant can connect to https://mcp.ungovr.org/mcp today and
call eight tools against the UnGovr Atlas, including one that files a service request
on a resident's behalf. Coverage is the honest limit. The tool set currently
reaches counties, state agencies, joint powers authorities and regulated
utilities. Cities and special districts are served by the website and the data
API but are not yet reachable through the tools, so the streetlight in Fig. 1 is
a question the UnGovr Atlas can answer and the MCP tools cannot. We are fixing that, and
we would rather you read it here than discover it.
| Tool | What it does |
|---|---|
search_entities | Find governments by country, region, name or type |
get_entity | One government in full, with its parent, boundary and contacts |
list_countries | The full country list, with counts and coverage |
get_product_coverage | Where a given UnGovr product is available |
list_service_types | What can be reported or requested in a place |
submit_service_request | File a report, such as that streetlight, on someone's behalf |
search_cgj_reports | Search California civil grand jury investigations |
get_cgj_report | One report in full, with findings and responses |
Where the rules get written
A public-interest voice in the standards, while they are still soft.
How agents talk to government is being settled now, in working groups, by the parties who show up. Almost all of them are companies. UnGovr is an Associate Member of the Agentic AI Foundation, a Linux Foundation initiative behind the Model Context Protocol, and of The Linux Foundation itself. We are there for one reason: to argue that the map of government belongs in the commons, and to publish ours that way so the argument has something behind it.
