MapLark OSM Features API

Official Python client for the MapLark OSM Features API (GeoJSON, FlatGeobuf, GeoParquet, CSV).

Contents

• Python SDK
• Quick start
• Basic API usage
  • Create a client
  • Query OSM features
  • Auto-pagination and bbox tiling
  • Async client
  • Convenience helpers
  • Cost and usage
• AI
  • MCP Server
  • Cursor / Claude Desktop
  • Typical AI questions
• Places and routes
  • Places search
  • Nearby
  • Place details
  • Opening hours
  • X near Y
  • Walk and bike routes
• CLI
• Example apps

Query OpenStreetMap features such as buildings, streets, and Points of Interest easily. Search for OSM features by bounding box, tags, and geometry shape and get GeoJSON back within less than 250ms (dependent on query size). No converting between formats manually. The API keeps OpenStreetMap semantics intact, like tags and ways, and returns GeoJSON Features you can feed straight into Leaflet, MapLibre, OpenLayers, or any geospatial toolchain. It is backed by postgis with tiered API keys and rate limiting to keep noisy neighbours out to give you predictable latency for real traffic.

The postgis translation layer is very simple:

• node - Point
• way - LineString or Polygon
• relation - MultiPolygon or grouped geometries

You filter with the same tags mappers already use (amenity=cafe, building=yes, and so on). Knowledge from OSM, Overpass, and tagging docs transfers immediately.

To narrow down between "open ways" and "closed ways", use the way_shape parameter:

• way_shape=line - open ways (roads, paths, rivers) or line-shaped relations (routes, boundaries)
• way_shape=polygon - closed ways (buildings, parks) or multipolygon relations.
• way_shape=all - both shapes (default when way_shape is omitted).

For example, to get all buildings in an area:

type=way & tags=building

This is the equivalent of the Overpass query way[building].

Read the full API reference here https://maplark.com/developer.

Python SDK

This client library comes with auto-pagination, bbox tiling (enables larger bbox queries), retry/backoff, pandas/geopandas output, async support, convenience methods for common OSM layers (buildings, amenities, bike roads, and so on), a stdio MCP server for Claude / Cursor / Custom agents, and places/routes methods for opening hours and walk/bike routing.

pip install osmfeatures
pip install "osmfeatures[geo]"   # pandas / geopandas / shapely support
pip install "osmfeatures[mcp]"   # stdio MCP server for Claude / Cursor

Official client for the MapLark OSM Features API (GeoJSON, FlatGeobuf, GeoParquet, CSV). The SDK talks to api.maplark.com by default.

Quick start

from osmfeatures import OSMFeaturesClient

with OSMFeaturesClient(api_key="sk-...") as client:
    fc = client.query(bbox="18.06,59.32,18.09,59.34", tags=["building"])
    print(len(fc.features), "buildings found")

Basic API usage

Create a client

from osmfeatures import OSMFeaturesClient

client = OSMFeaturesClient(api_key="sk-...")

You can use the client directly and close it when done, or use a context manager:

from osmfeatures import OSMFeaturesClient

with OSMFeaturesClient(api_key="sk-...") as client:
    ...

Query OSM features

query() fetches a single page:

fc = client.query(
    bbox="18.063,59.322,18.082,59.332",
    type="way",
    way_shape="line",
    tags=["highway=cycleway"],
    limit=500,
)

for feature in fc.features:
    print(feature["id"], feature["geometry"]["type"], feature.tags)

Common filters:

• bbox="min_lon,min_lat,max_lon,max_lat"
• location="lat,lng" with radius in metres
• tags=["amenity=restaurant"] (AND)
• or_tags=["bicycle=yes", "bicycle=designated"] (OR)
• not_tags=["access=private"] (exclude)
• type="node" | "way" | "relation"
• way_shape="polygon" | "line" | "all" (omit = both shapes; all also means both)
• clip_geometry=True | False (omit for the API default True; set False to keep full geometry outside bbox)
• cursor (pagination; use SDK meta.next_cursor from previous page, sourced from X-Next-Cursor)

Auto-pagination and bbox tiling

Use query_all() to fetch all pages and deduplicate by OSM feature id. By default it splits the bbox into 2 tiles (power of 2) so large areas use more requests; pass bbox_tiles=1 to disable, or raise it (4, 8, …) for bigger areas:

all_restaurants = client.query_all(
    bbox="18.063,59.322,18.082,59.332",
    tags="amenity=restaurant",
    max_features=55_000,  # client total cap; pass None for no cap
)

print(all_restaurants.meta.returned)

Async client

Async methods mirror the sync API (query_async, query_all_async):

import asyncio
from osmfeatures import AsyncOSMFeaturesClient


async def main() -> None:
    async with AsyncOSMFeaturesClient(api_key="sk-...") as client:
        fc = await client.query_async(
            bbox="18.06,59.32,18.09,59.34",
            tags=["building"],
        )
        print(len(fc.features))


asyncio.run(main())

Convenience helpers

For common datasets, use convenience methods built on top of query_all():

from osmfeatures import OSMFeaturesClient, get_buildings, get_restaurants

with OSMFeaturesClient(api_key="sk-...") as client:
    buildings = get_buildings(client, bbox="18.063,59.322,18.082,59.332")
    restaurants = get_restaurants(client, bbox="18.063,59.322,18.082,59.332")
    print(len(buildings.features), len(restaurants.features))

Cost and usage

estimate = client.estimate_cost(
    bbox="18.063,59.322,18.082,59.332",
    tags=["building"],
)
print("estimated credits:", estimate.estimated_credits)

usage = client.usage()
print("Usage:", usage)

AI

Add real geospatial intelligence to your Artificial Intelligence agents. Use Maplark in Claude Code or build new custom agentic apps.

MCP Server

The Maplark MCP Server is the agent surface. Your LLM is the planner: it chooses OSM tags, a bbox or location+radius, a time, a travel mode, and the next tool. The tools compute metres, ranks, opening-hours status, and walk/bike paths. You do not compute haversine, parse opening_hours strings, or invent coordinates.

Results come back as summaries (ids, names, OSM tags, lon/lat scalars, distance_m, openNow) plus a collection_id. They never include GeoJSON coordinate arrays. Call preview_map(collection_id) to draw: a local page loads OpenFreeMap (Liberty) in MapLibre and fetches GeoJSON from localhost, so coordinates never enter the model. Call export_geojson only when the user asked for a raw file: it writes GeoJSON to disk and returns a path, not coordinates.

Named places go through geocode (Nominatim until MapLark /v1/geocode is public). For a large bbox, call query_all with bbox_tiles (power of 2; 1 disables tiling), not page query by hand. MCP query_all defaults to max_features=10000 (raise it if has_more); the SDK default remains 55_000. Do not invent places_near_to or places_open_after. The server ships these rules as instructions from src/osmfeatures/mcp/AGENT_INSTRUCTIONS.md.

Tools

• Places: places_search, places_nearby, places_details
• Geocode: geocode (Nominatim interim)
• Routes: routes_isochrone, routes_path, routes_optimized_path
• Generic OSM: query (one page), query_all (tiled pages)
• Local (no HTTP): nearest_within, filter_open, point_in_polygon, points_in_polygon
• Draw / export: preview_map, export_geojson

Run MCP server manually

pip install "osmfeatures[mcp]"
export MAPLARK_API_KEY="sk-..."
osmfeatures mcp

Cursor / Claude Desktop

First install uv for one-click server start.

{
  "mcpServers": {
    "maplark": {
      "command": "uvx",
      "args": ["--from", "osmfeatures[mcp]", "osmfeatures", "mcp"],
      "env": { "MAPLARK_API_KEY": "YOUR_KEY" }
    }
  }
}

Planner rules: you pick tags, bbox or location+radius, budgets, openNow/asOf, and the next tool. Code computes metres, ranks, network paths, and opening-hours status.

Typical AI questions

Prompt	MCP tools
"Cafes near me"	places_nearby or places_search with location+radius / bbox (hours only for staffed amenities)
"Waste baskets / EV chargers / hotels in the area"	places_search with no as_of / open_now
"Restaurants within 150 m of a station"	two places_search, then nearest_within
"Bars open at 20:00"	places_search with as_of (no open_now so closed hits stay), then filter_open; if empty, retry with no hours
"Cafes within a 10-minute bike ride"	routes_isochrone + places_search in a covering radius + points_in_polygon
"A walking bar crawl in Stockholm" / "cafes on a tour of Gamla Stan"	places_search (now, or retry as_of if all closed; drop hours if still empty), then filter_open + routes_optimized_path (loop=true)
"Walk from my hotel to the cafe, then the office"	routes_path with those stops in listed order
"Suggest a walk to a bar, a restaurant, and a cafe, no particular order"	routes_optimized_path with loop=false
"Is the office a 20-minute walk from the apartment?"	routes_isochrone from A, point_in_polygon for B
"Show this on a map"	preview_map(collection_id) after a search or route

The same operations exist on OSMFeaturesClient when you are not going through an LLM (see Places and routes). Full HTTP reference: https://maplark.com/developer.

Places and routes

query() is the generic OSM layer: buildings, roads, park polygons, any tag and geometry shape. Places and routes are the place and mobility layer on top of the same data. You pick OSM tags, an area, a time, and a travel mode. The API returns coordinates, opening-hours status, straight-line ranks, and walk/bike geometry. You do not compute metres or parse opening_hours strings yourself.

These are the same operations as the MCP tools; call them on OSMFeaturesClient when you are not going through an LLM.

Runnable Python chains live in tests/example_apps/test_geo_agent.py. Full HTTP reference: https://maplark.com/developer.

Places search

places_search() finds places in a bounding box or a location plus radius (not both). Optional tags (AND) and or_tags (OR) use the same OSM filters as query(). Omit limit to use the API default (100, max 10_000).

cafes = client.places_search(
    location={"lat": 59.316, "lng": 18.075},
    radius=800,
    or_tags=["amenity=cafe"],
    open_now=True,
    as_of="2026-08-10T18:00:00+02:00",
)
print(len(cafes["features"]), "open cafes")
print(cafes["metadata"]["evaluated_at"])

Response is a GeoJSON FeatureCollection plus metadata.evaluated_at (UTC instant used for hours).

Nearby (ranked from a point)

places_nearby() answers "X near this point". It requires tags or or_tags. Results are ranked by straight-line spheroid distance, nearest first. Omit radius / limit to use the API defaults (1000 m / 100).

nearby = client.places_nearby(
    location={"lat": 59.316, "lng": 18.075},
    or_tags=["amenity=cafe"],
    limit=5,
    open_now=True,
    as_of="2026-08-10T18:00:00+02:00",
)
for item in nearby["items"]:
    print(item["distance_m"], item["feature"]["id"])

Response: {status, items: [{feature, distance_m}], estimated_units, evaluated_at}.

Place details

places_details() loads one place by the id that search or nearby returned (node/123). You can pass that string, or osm_type plus osm_id. Missing or non-place ids return HTTP 404.

details = client.places_details(cafes["features"][0]["id"])
# same as: client.places_details("node", 123)
print(details["feature"]["properties"]["tags"])
print(details["timezone"], details["evaluated_at"])

Response: {status, feature, estimated_units, evaluated_at, timezone}. Hours are annotated at request time in the place's IANA zone (from its coordinates).

Opening hours

Every place feature includes properties.openNow (true / false) when hours are evaluable. The field is omitted when hours are missing or unparseable.

Hours use each place's IANA timezone from its coordinates. There is no request timezone field.

• open_now=True keeps only known-open places. Missing or unparseable OSM opening_hours are dropped (same idea as Google Places openNow).
• as_of is the evaluation instant (default: now). A value with an offset (Z or +02:00) is an absolute instant. A naive value (2026-08-10T20:00:00, no offset) is that local clock at the search location or bbox center.
• Passing as_of or open_now also requires an OSM opening_hours tag, so untagged POIs do not fill the page.
• Closed places that have hours still return unless open_now is set.

"X near Y" (local join)

nearby ranks against one point. "Restaurants within 150 m of a station" is two searches plus a local join. nearest_within does no HTTP.

from osmfeatures import nearest_within

bbox = "18.05,59.33,18.10,59.36"
restaurants = client.places_search(bbox=bbox, or_tags=["amenity=restaurant"])
stations = client.places_search(bbox=bbox, or_tags=["railway=station"])
pairs = nearest_within(restaurants, stations, max_distance_m=150, limit=20)

for pair in pairs:
    print(pair["distance_m"], pair["feature"]["id"], "near", pair["nearest"]["id"])

Each pair is {"feature": <primary>, "distance_m": <float>, "nearest": <secondary>}. The point comes from geometry when it is a Point, else properties.centroid. A feature with neither raises ValueError. Empty secondary returns []. Distances are spherical haversine (mean Earth radius 6371000 m). Fine at search limit (default 100).

Walk and bike routes

Routing follows the OSM walk or bicycle network (query-time Dijkstra on tiled highways). Omit travel_mode to use the API default (WALK), or pass "BICYCLE". Walk treats the graph as undirected (oneways ignored). Bicycle is directed and honors OSM oneway, oneway:bicycle, contraflow cycleways, and implied roundabout oneway. Car routing (DRIVE) is not available.

Duration budgets convert at about 5 km/h for walk (1.4 m/s) and 15 km/h for bicycle (4.2 m/s). Optional search_buffer_m widens the highway fetch corridor if a path cannot be formed in the default area.

Router endpoints return an OSRM-style status in a 200 body (not always HTTP 4xx):

• ok
• area_too_large_for_tier (no graph fetch)
• tile_too_dense
• start_unreachable / end_unreachable
• no_path_within_area

Always check status before reading geometry.

Isochrone. Reach polygon from origin. Provide exactly one of max_distance_m or duration_s. Geometry is a buffered union of reachable edges (city blocks stay holes).

origin = {"lon": 18.075, "lat": 59.316}
iso = client.routes_isochrone(origin=origin, duration_s=600, travel_mode="WALK")
if iso["status"] == "ok":
    print(iso["geometry"]["type"], iso["distance_m"], iso["duration_s"])

Path. Given-order walk or bike through 2 to 250 stops. Does not reorder stops or close a loop. Two stops is A to B. Three or more stitches legs and returns stop_distances_m. To walk a known sequence home, repeat home as the last stop. Unordered search hits belong on routes_optimized_path.

path = client.routes_path(
    stops=[origin, {"lon": 18.08, "lat": 59.318}],
    travel_mode="WALK",
)

Optimized path. Tour from start through unordered stops (nearest-neighbour + 2-opt). Do not put start in stops. loop=True (default) returns to start. loop=False is an open path that ends at the last ordered stop. Response includes ordered_stops (start first).

opt = client.routes_optimized_path(
    start=origin,
    stops=[{"lon": 18.08, "lat": 59.318}, {"lon": 18.07, "lat": 59.320}],
    loop=True,
    travel_mode="WALK",
)
print(opt["status"], opt.get("ordered_stops"), opt.get("distance_m"))

Points accept lon or lng. Places methods send {lat, lng}. Route methods send {lon, lat}.

Local helpers (no HTTP): nearest_within(primary, secondary, max_distance_m) for "X near Y", and point_in_geometry(lon, lat, geom) for isochrone containment.

CLI

If the package is installed, the CLI is available as osmfeatures:

export MAPLARK_API_KEY="sk-..."
osmfeatures query --bbox "18.063,59.322,18.082,59.332" --tags building --type way
osmfeatures query --bbox "18.063,59.322,18.082,59.332" --tags building --all-pages --bbox-tiles 4
osmfeatures mcp

Example apps

The repository includes runnable example-app tests in tests/example_apps/ showing end-to-end usage patterns against real OSM data.

• test_restaurant_guide.py: restaurant discovery list with names/cuisines and map coordinates.
• test_park_bench_finder.py: bench finder for park maps (amenity=bench).
• test_park_explorer.py: park browser with polygon boundaries, centroids, and area estimates.
• test_cycling_trails.py: unpaved cycling trail layer for MTB/gravel planning.
• test_city_cycling_infrastructure.py: city cycling overlay combining cycleways and bike lanes.
• test_lakeside_ice_cream_hunt.py: nearest ice cream shops to waterfront edges.
• test_pedestrian_shortest_path.py: shortest walking route via graph + Dijkstra.
• test_pedestrian_wavefront_bfs.py: hop-based accessibility rings via BFS.
• test_bike_path_dijkstra_liljeholmen_to_djurgarden.py: tiled corridor bike routing from Liljeholmen to Djurgarden.
• test_geometry_filters.py: zoom + area/length filters for large buildings and long roads.
• test_geo_agent.py: Python places/routes chains (bar crawl, bike parks, isochrone filter/compare/coverage, client-side open-at-clock).

Run all example apps:

pytest tests/example_apps -v

Run one example app:

pytest tests/example_apps/test_restaurant_guide.py -v