Built for AI agents, not retrofitted.

Forward geocoding: name, label, position and address parts per candidate.

Coordinates to the nearest places: addresses, POIs and localities, nearest first with distance in metres. The inverse of geocode. Pass a heading and results come back described from where the user is standing rather than as coordinates.

Nearest-first search around a point, by category (cafes, fuel, EV charging, parking) or by name and brand, with real distances. Pass heading_deg and every result is phrased egocentrically, “ahead and slightly to your right, about 80 metres”, with a signed relative bearing; add fov_deg to keep only what is in front of the user. Without a heading it falls back to cardinals, so a client with no compass still gets something it can read aloud.

What is over there, and whether you could see it. Give it a position and a bearing and it returns the places in that cone, each with its distance, how far off dead ahead it sits, and an honest verdict: clear, occluded with the obstruction named, or unknown when nothing could be checked. Visibility is modelled from the terrain and OpenStreetMap building heights, never observed, and the response says so.

What a journey passes, in sentences ready to read aloud: the named rivers it crosses, the road it runs on and for how far, the settlements it goes through, the landscapes it enters and how high the road climbs. None of it is in a route object, so it is computed by intersecting the route geometry with the basemap and the terrain model. The silence budget is the design, and the coverage block separates “nothing there” from “nothing read”.

Three to five short lines an hour about the ground the car is on, each at the point it belongs. Every sentence was written from a public-domain plaque inscription and reviewed by a person before it shipped, so no model runs while anybody is driving. It covers one corridor on purpose, and every other route answers with no lines and a census of what was out of reach.

Check places and itineraries a model mentioned against the map before you act on them. Each claim comes back verified, contradicted or unverified, never a boolean, and a missing match is unverified rather than a denial. Pass a locality with every claim: a bare name is searched worldwide, so a short name can verify against a same-named place in another country.

Many-to-many travel-time and distance matrix; null cells mark unreachable pairs.

Stops from your own places dataset along a route, ranked by honest detour cost: real added driving time, measured by the engine.

Isochrone rings: where you can reach inside a chosen time budget on foot, by bike or by car.

Turn a list of stops into a navigable multi-stop route with per-stop arrival times; optionally optimises the order.

Terrain height for a list of points, or a resampled profile along a route. Null wherever the terrain set has no coverage, never a guess.

Snap a recorded GPS trace to the roads it actually travelled, under the costing it was travelled with, so a walk is not matched onto the carriageway.

The canonical category tokens and their colloquial aliases. Read it before guessing: an unknown token matches nothing and returns empty rather than erroring.

Metres between two coordinates, as the crow flies. Not a driving distance: use route or matrix for that.

Initial bearing between two coordinates, degrees clockwise from true north.

The coordinate reached by travelling a distance on a bearing. The inverse of geo_distance plus geo_bearing.

Is this coordinate inside the zone? Delivery areas, catchments, clean-air and congestion zones. On the boundary counts as outside.

The axis-aligned box enclosing a set of coordinates, for fitting a map view to a set of stops.

The centroid of a set of coordinates: pick a depot, centre a map.

Total length of a polyline in metres. Measures the line you supply, not a driven route.

Area in square metres inside a ring. Always positive, so winding direction cannot change the answer.

Douglas-Peucker simplification. The tolerance is in degrees, not metres, and endpoints are always kept.

The closest position on a polyline to a coordinate, plus the distance to it. Answers “how far is this address from the route?”

Does a dataset's declared CRS actually describe its own coordinates? Catches swapped axes, degrees labelled as metres and a mislabelled projection before anything is drawn. A check, never a reprojection.

Queue a map correction for review: location, category, evidence. Never an automatic OpenStreetMap edit.

Send a structured end-of-integration retro so the friction you hit gets fixed. Only the named fields travel, never your conversation, prompts or code, and only with the developer's approval.

What this key has spent, per day and per endpoint, against its quota and prepaid balance. Reading it is free, so an agent can decide mid-task whether to keep going.

Incidents in force near a point, in a box or along a route, each with the clip it was verified on, the confidence and when it lifts. An empty answer is no camera-verified incident, never a clear road.

An operator's short-lived closure or speed on named roads, admin only, and only after a person approves that exact closure.