Maps for AI agents: the 2026 landscape

Every mapping platform now ships a tool server for agents. Most agents still answer "where should I meet them" with invented coordinates. This is a survey of what the major map and location tools ship today, why that is not enough, and a checklist to apply to any tool you wire into an agent.

What changed this year

The Model Context Protocol (MCP) turned "call a map API from an agent" into a solved plumbing problem: point an agent at a server, let it discover tools at runtime. Tracking catalogs now list over 100 geospatial MCP servers: 47 commercial, 41 published by the data vendors themselves. Google, Mapbox, TomTom, Esri, CARTO, Felt, Cesium, and Baidu all ship official servers. The plumbing is done. What has not caught up is what comes back down the pipe.

At the same time a newer, smaller group of tools started marketing on a different axis: not more data, but less of it - compact answers sized for a context window instead of raw GeoJSON dumps. That second wave is where the interesting engineering is, and it is where this survey focuses.

Four failure modes that still break agents

An agent with a map fails in particular, repeatable ways. Any "maps for agents" tool should be judged against all four.

  1. Hallucinated coordinates. A model asked to reason over raw lat/lon lists will, with nonzero probability, invent them. This is not a prompt-engineering problem; it is what happens when the format is something a model was barely trained to reason in.
  2. Silent truncation. A fetch of "everything near here" returns hundreds of kilobytes, the agent framework cuts it to fit, and the map the model reads is a random prefix that looks complete. The model cannot tell not there from not shown. It confidently reports the pharmacy that got truncated away does not exist.
  3. Fictional routing. Routing APIs sometimes return confident nonsense - impossible distances, routes through water. In our own benchmark work, the public demo router answered Ok with a walk that covered a straight-line half-kilometre in 40 seconds. Nothing in the response said "check me", so nothing downstream did.
  4. Invented hours. A place with no recorded hours is not a closed place, but outputs that render "unknown" as "closed" (or as open) push an agent to send a person on a wasted trip - the exact failure an honest refusal is supposed to prevent.

Who ships what

The platforms came for integration; the new wave came for context budgets. Fair summary as of October 2026:

ToolShape of an answerNotes
Mapbox MCPraw JSON / GeoJSON payloadsOfficial, hosted endpoint, full API coverage. Metered API billing. Answer size is your problem.
Google Maps Grounding LiteAI place summaries + payloadsOfficial, hosted. Place summaries are model-generated en route - convenient, and a new thing to trust.
TomTom Agent Toolkit53 tools + shared state layerSmart design: full geometry lives in state, the model reads one-line summaries. Front-end oriented.
PlaceRootranked compact answers from OvertureKeyless, 48 tools, real graph routing. The strongest "answers, not dumps" implementation outside the platforms.
Footsteptoken-lean JSONHosted MCP, 15 tools, geometry opt-in. Freemium.
GeoContextone-call context bundlePOI + neighbourhood profile + timezone per request. Claims an 80% tool-call overhead cut.
xMap / RSMaps / Foursquarecommercial datasets via toolsLarge licensed POI/mobility corpora, agent-facing schemas.
supercartotoken-budgeted topological graphOpen source. The map is a walkable graph, cut to a budget re-measured on the emitted output, with every removal recorded. Public re-derivable benchmark.

Two structural camps emerge. Vendors host their data and keep the API key - convenient, metered, and the answer shape is decided by the vendor's API, not by what a model reasons over well. Open-data tooling runs locally - Overpass, Overture, self-hosted PMTiles - where the shape is a choice the tool author makes. Most of the open tools still return flat ranked lists. Flat lists answer "what is here". They do not answer "how do the things here connect", which is what an agent needs when the next step is moving a person.

The checklist

Before wiring any map tool into an agent, ask for these five things. If a tool cannot answer one of them, that is the data point.

  1. Is the token budget a budget? Not "compact by design" - a number that is re-measured on the actual emitted output, guaranteed, with a fallback when it cannot be met.
  2. Does the map survive its own cuts? After truncation, is the remaining area still traversable - can the agent route between what remains? A trimmed map that strands its landmarks is worse than a smaller complete one.
  3. Can the agent tell not there from not shown? Truncation, missing data, and a closed/open-at-noon-checked-against-real-time distinction must be recorded in the output, not on the website.
  4. Is there a public score? Not marketing claims - tasks, models, pass rules, and raw answer archives that a third party can re-derive without model calls. If the numbers cannot be checked, they are a mood.
  5. What happens when the source is down? Zeroes and empty strings read as "no rain" and "nothing here". A degraded answer must announce itself as degraded.

The gap that was next: transit

When this survey was first written, transit schedules were the clearest missing capability: the routine agent task - "can you be somewhere by 9?" - had no answer from any map tool for agents, and tracking catalogs showed no agent-facing GTFS server. That gap has now started to close from both directions. Dedicated transit-only MCP servers exist (MTA, Toronto, and worldwide feed catalogs), and supercarto itself now ships get_transit: stops, lines and first/last departures from any static GTFS feed, inside the same token-budgeted maplet as the streets, terrain, weather and daylight of one place - not as a separate server an agent must already know to consult.

Where supercarto stands

supercarto takes the position that an agent should read something built for reasoning, not rendered for humans: a topological graph with distances in metres and compass directions, cut to a token budget that is re-measured on the delivered output, with every drop recorded. In the public benchmark - same questions, same models, only the context changes - the compiled graph scores 38/47 against 15/44 for raw GeoJSON and 0/47 for no map. Every number is re-derivable from committed answer archives without a model call. Indoor multi-storey graphs, keyless weather, honest degraded paths, and GTFS transit schedules ship in the core.