GeoWire

Give any AI agent real-world location intelligence in 5 minutes — no API key required.

One interface for place search, directions, and area/market analysis across every map provider.

GeoWire is an open-source geo intelligence gateway that sits between AI agents and map/place data providers (OpenStreetMap, Google, Yelp, Foursquare, US Census, your own data) and exposes them through a single MCP server, REST API, and SDK. One interface for place search, geocoding, directions & distance matrices, and commercial-area analysis (density, ratings, demographics) — with provider fallback, multi-provider merge + dedup, cost budgets, and a policy engine that enforces each provider's caching/attribution terms.

Published on npm — MCP · REST · CLI · SDK all functional. 10 providers, 8 operations.

Honest by design: OpenStreetMap (the zero-key default) is a great geocoder — strong on place names, addresses, and landmarks — but thin on category words ("coffee", "pharmacy"), ratings, and hours. For US/Western commercial data, add a Google, Yelp, or Foursquare key (BYOK) — GeoWire merges them and tells you which source every field came from. Directions run key-free on OpenStreetMap routing (OSRM); US demographics come from the Census (free key).

Contents: Why · Quickstart · MCP tools · REST · Anatomy of a response · Config · Providers · Recipes & examples · Roadmap · Architecture

Why GeoWire?

Not a Google replacement — it uses Google. The thing no single provider can do: merge your own store data + Google + OSM into one deduped record, with per-field provenance (your name is authoritative, Google adds ratings, OSM adds coordinates). Real run below:

More than search — location intelligence for agents

Once your agent can find places, it can reason about areas. One analyze_area call turns a point + radius into a commercial-district read — category density, competition, the rating landscape, and (in the US) demographics:

// "Is this a good block for a new café?" — SF, 1 km radius
{ "center": { "latitude": 37.7749, "longitude": -122.4194 }, "radiusMeters": 1000,
  "categories": ["cafe", "restaurant", "bar"] }
// → per-category counts + density/km², avg rating (Google/Yelp), price mix,
//   an activity proxy (Yelp review volume), and Census demographics for the tract.

Same gateway also does directions & distance matrices and travel-time isochrones — key-free via OSRM — so an agent can rank candidates by drive time and answer catchment questions like "what's within a 15-minute drive of this address?" See Recipes.

Related MCP server: NextBillion.ai MCP Server

Quickstart

1. MCP (Claude Desktop / Cursor) — 30 seconds

Add this to your MCP client config (e.g. Claude Desktop claude_desktop_config.json):

{
  "mcpServers": {
    "geowire": { "command": "npx", "args": ["-y", "@geowirehq/mcp"] }
  }
}

Then ask: "Where is the Golden Gate Bridge?", "Find coffee within 2 km of 37.7749, -122.4194 (San Francisco).", or "How do I drive from downtown SF to Fisherman's Wharf?" Works with zero API keys — OpenStreetMap + OSRM routing are the defaults. Add "env": { "GOOGLE_MAPS_API_KEY": "..." } (or YELP_API_KEY) for ratings, reviews, and hours, and "CENSUS_API_KEY" for area demographics. See more MCP client configs.

2. CLI — one-shot search & server

npx @geowirehq/cli search "Golden Gate Bridge"    # terminal search with a results table
npx @geowirehq/cli search "coffee" --near 37.7749,-122.4194 --radius 2000   # near a coordinate
npx @geowirehq/cli reverse 37.8199,-122.4783      # coordinate → nearest place
npx @geowirehq/cli route 37.7749,-122.4194 37.8083,-122.4156   # driving directions (no key, OSRM)
npx @geowirehq/cli get google:ChIJ...             # one place by reference (getPlace-capable provider)
npx @geowirehq/cli                                # start the REST + MCP server (zero-config)
npx @geowirehq/cli init                           # interactive setup wizard (.env + config)
npx @geowirehq/cli test                           # check provider connections

Add --json to any command for the full response (results + provenance meta).

3. Docker — self-hosted server

docker run -p 4980:4980 geowire/geowire
# then:
curl -X POST http://localhost:4980/v1/places/search \
  -H 'content-type: application/json' \
  -d '{"query":"coffee","near":{"latitude":37.7749,"longitude":-122.4194},"radiusMeters":2000}'

Or with docker compose up (see docker-compose.yml). API docs at /docs.

4. SDK (embedded)

import { createGeoWire } from "@geowirehq/core";
import { createNominatimProvider } from "@geowirehq/provider-nominatim";

const geo = createGeoWire({ providers: [createNominatimProvider()] });
const { results, meta } = await geo.searchPlaces({
  query: "coffee",
  near: { latitude: 37.7749, longitude: -122.4194 },
  radiusMeters: 2000,
});

Full embedded-SDK guide: examples/typescript-sdk.md.

MCP tools

Every response includes both a human-readable summary and structuredContent (schema-valid JSON).

REST endpoints

Optional Bearer auth: set GEOWIRE_API_KEYS=key1,key2.

Anatomy of a response

No black box. Every response carries a meta block: which providers were used / skipped / failed (and why), dedup counts, cache status, estimated cost, and per-field sourcing — so you always know where each value came from.

{
  "results": [{
    "id": "gwp_CvWvRZrFtegkJPxP9CW0",
    "name": "Golden Gate Bridge",
    "location": { "latitude": 37.8199286, "longitude": -122.4782551 },
    "sources": [{
      "provider": "nominatim",
      "providerPlaceId": "way/27385590",
      "fields": ["name", "location", "categories", "address"]   // ← what this source contributed
    }],
    "attributions": ["© OpenStreetMap contributors"]
  }],
  "meta": {
    "providersUsed":   [{ "provider": "nominatim", "resultCount": 1, "latencyMs": 2449 }],
    "providersSkipped": [],   // e.g. { provider: "google", reason: "MISSING_CREDENTIALS" | "QUOTA_EXCEEDED" }
    "providersFailed":  [],   // e.g. { provider: "google", reason: "TIMEOUT" }
    "strategy": "first-success",
    "cache": { "hit": false }
    // merging adds:  "dedup": { "before": 3, "after": 1 }
    // paid provider: "estimatedCostUSD": 0.032
  }
}

After a merge, sources[].fields shows (say) the phone came from Google while the coordinates came from OSM. Walkthrough: docs/recipes.md.

Configuration (optional — everything works without it)

geowire.config.yaml:

providers:
  nominatim: { enabled: true }                       # default ON, no key (search/geocode)
  osrm:      { enabled: true }                        # default ON, no key (directions)
  google:    { enabled: true, apiKey: ${GOOGLE_MAPS_API_KEY} }
  yelp:      { enabled: true }                        # env YELP_API_KEY (US/Western business & reviews)
  census:    { enabled: true }                        # env CENSUS_API_KEY (US demographics, free)
  kakao:     { enabled: true }                        # env KAKAO_REST_API_KEY (KR)
  internal:  { enabled: true, source: ./my-places.csv, priority: 100 }
routing:
  defaultStrategy: merge          # first-success | merge | cost-aware | weighted | fastest
  providerWeights:                # for `weighted`: order by priority·cost·coverage
    priority: 0.5
    cost: 0.3
    coverage: 0.2
budget:
  perRequestMaxUSD: 0.10          # over-budget paid providers are skipped, free ones used

Keys come from the environment (${VAR}), never committed in plaintext.

Providers

Regional providers make Korea (Kakao/Naver) and China (Baidu) coverage first-class where OSM is thin and Google has gaps — Baidu returns BD-09 coordinates, which GeoWire converts to WGS84 automatically. Merge them all + your own store data into one deduped record.

Provider roles — each provider does what it's best at

Providers aren't interchangeable; they're complementary. When merge combines duplicates, GeoWire doesn't just pick the highest-priority provider's whole record — it sources each field from the provider that's authoritative for it. Every provider declares its strengths in its manifest (fieldAuthority), so one merged place can carry OSM's coordinates, Google's reviews, and Kakao's local name at once:

sources[].fields in every response records which provider contributed which field. This is the "Stripe for Maps" idea in code: you get one clean place record, and each part of it comes from whoever knows it best. (Reviews/photos are provider originals — the policy engine enforces each provider's storage terms; Google originals aren't cached.)

Want another provider? See CONTRIBUTING.md — "Write a provider in 30 minutes".

Recipes & examples

• docs/recipes.md — end-to-end recipes: near+radius search, merge + dedup, cost budgets, country routing, your own CSV, self-host.

• examples/mcp-clients.md — configs for Claude Desktop/Code, Cursor, Cline, VS Code, Windsurf.

• examples/typescript-sdk.md — embed the SDK.

• examples/llm-tool-use.md — raw OpenAI / Anthropic function calling. Also LangChain · Vercel AI SDK.

Roadmap

v0.1 is deliberately "It works" scope. Honest about what's not in it yet:

Architecture

AI agent / app
   │  MCP · REST · SDK
   ▼
GeoWire core  ── pipeline: plan → execute → normalize → dedup → rank → policy → cache
   │  GeoProvider contract
   ▼
providers: nominatim · osrm · google · yelp · foursquare · census · kakao · naver · baidu · internal · (community)

Monorepo packages: schema · provider-sdk · provider-testkit · core · providers/* · mcp · apps/server · cli.

Documentation

• Recipes / cookbook — task-oriented, copy-pasteable

• Examples — MCP clients, SDK, LangChain, AI SDK, tool use

• Contributing + write a provider

• System design

License

Apache-2.0. GeoWire's code license is separate from the terms of third-party map/place data providers — usage of Google, Mapbox, HERE, Kakao, Naver, etc. is governed by each provider's own terms. OSM data is under ODbL; GeoWire's policy engine enforces attribution and caching limits per provider.