topoloop
🏆 Beats Google Places on real agent queries →

The places API built for AI agents

Semantic search over 49 million places at 1% of Google's price — licensed to cache, store, and train on. Your agent can finally afford to think.

Start building — free $ claude mcp add topoloop
10,000 free calls/mo No card required Hard spend caps
Live demo agent → topoloop · POST /places/search
// your agent, mid-loop:
200 OK · 236 ms · top of 5
© Mapbox © OSM
Yari-ga-take
槍ヶ岳 · Hida Mountains, JP
49M
places indexed, planet-wide
100+
languages understood
300ms
typical p50, warm
30d
full-planet refresh from OSM
Platform

Every geo capability an agent needs.
Nothing a human search box wants.

Purpose-built for the two things agents actually do with geography: find the place a human meant, and resolve the places a document mentions.

Semantic place search

Agents don't typo — they approximate. "The golden temple in Kyoto," a stripped romanization, an epithet from a blog post — an embedding model fine-tuned on real agent traces resolves the place the model means, not the string it sent.

Built for loops

Responses are candidate pools with scores, not a single take-it-or-leave-it answer. Your LLM disambiguates; you validate its pick by reference. The pattern behind pipelines that resolved 150k+ places from photos.

MCP-native

One line adds topoloop to Claude, Cursor, or any MCP client. Tool schemas designed for function-calling — with OpenAI, LangChain, and Vercel AI SDK bindings that behave identically.

Fame-aware nearby

"What's around here?" shouldn't return forty lampposts. Prominence signals mined from the open web surface the temple over the vending machine — world landmarks are guaranteed a seat within radius.

The enrichment layer

Verified official websites, live hours, services, and one-line descriptions — independently sourced, keyed to every place, licensable as a clean sidecar. The columns Google won't sell you.

Open data, no lock-in

Built on OpenStreetMap, re-ingested from the raw planet file monthly. Stable tl_ IDs with OSM crosswalk. Self-host bundles when you need the index inside your own perimeter.

How it works

Anatomy of a resolution loop

The route your agent walks — four steps, sub-second round trip. Works identically for a chat agent geocoding "that ramen place near the station" and a batch pipeline resolving a million blog posts.

1
Search

Ask in any words

Your agent calls search or nearest with whatever it has — a name, a phrase, a coordinate, another language.

2
Candidates

Get a ranked pool

Scored candidates with names, coords, fame, and context — sized for a model to read, not a human to scroll.

3
Disambiguate

Your LLM picks

The model chooses with full context of its task. Ambiguity is resolved where the intelligence is — in your loop.

4
Validate

Lock the reference

The pick is validated against the pool and returned as a stable tl_ ID. Hallucinated places can't survive this step.

Quickstart

Five minutes to first fix

From zero to a production tool-call before your coffee cools. Pick your trailhead:

MCP server for Claude, Cursor, Windsurf
REST — two endpoints, obvious JSON
Python & TypeScript SDKs, typed
LangChain tool & Vercel AI SDK bindings
# add topoloop to Claude Code
$ claude mcp add topoloop \
    --url https://api.topoloop.com/mcp \
    --header "Authorization: Bearer $TOPOLOOP_KEY"

# your agent now has:
search_places  — semantic search, any language
nearby_places  — fame-aware "what's around here"
place_details  — hours, website, enrichment
Benchmark

Surveyed against real agent queries

Our eval set is 1,000+ tool-calls made by production LLM agents — cross-language, misspelled, half-remembered. Not benchmark-friendly strings. The full public harness (vs. Google, Foursquare, Mapbox, LocationIQ) ships as a repo you can re-run.

System Recall@1 Recall@5 Recall@10
topoloop (retrieval + reranker) 0.6660.8600.908
topoloop (semantic only) 0.5820.8140.865
Google · Foursquare · Mapbox · LocationIQ public benchmark repo — in the field, publishing soon
1,037 held-out (query, place) pairs from live agent traces · measured June 2026 · methodology & harness will be open source
Actually buildable

Location data you can build a business on

Affordable on day one. A real path to production. And no rug-pull priced into the terms — cache it, store it, train on it, render it on any map you like. If you outgrow the API, the self-host bundle means you can never be stranded.

What you can do with results
topoloop
Google
Cache responses indefinitely
✓
✗
Store place IDs in your DB
✓
30 days
Display on any map (MapLibre, Mapbox…)
✓
✗
Use in training & evals
✓
✗
Bulk export your resolved data
✓
✗
Self-host the full stack
✓
✗
Pricing

Priced for loops, not sessions

Agents are chatty. Google charges $32 per thousand searches and Mapbox bills "sessions" at up to $11.50 — pricing built for humans typing in boxes. Agents run loops, so we price for loops. Hard caps on every tier — a runaway loop hits a wall, never your card. Volume pricing published all the way up: no "contact sales" cliff between you and scale.

Live now
Base Camp
$0 /mo
  • 10,000 calls / month
  • 2 requests / second
  • Search + nearby + details
  • MCP server access
  • Community support
Start free
Coming soon
Trailhead
$29 /mo
  • 75,000 calls / month
  • Enrichment fields included
  • Batch endpoints
  • Email support
Coming soon
Coming soon
Expedition
$199 /mo
  • 600,000 calls / month
  • Entity-resolution pipelines
  • 99.9% uptime SLA
  • Priority support
Coming soon
Summit / Self-host
Custom
  • Run the full stack inside your perimeter
  • Regional or planet bundles
  • Monthly index + model refresh
  • Data licensing available
Talk to us

The free tier is live today — paid tiers are coming soon, and the prices above are what they'll launch at. Past 600k calls: published volume rates from $0.30 / 1k, down to $0.25 / 1k above 5M — your bill grows smoothly from $0 to scale without a single sales call. True pay-per-call: no reserved capacity, no per-second throttles.

Your agent already knows what.
Give it where.

10,000 free calls a month. One line of MCP. The planet, refreshed monthly. Bring the loop — we'll bring the map.

Get your API key
36°20′31″N 137°38′51″E · CONTOURS: HIDA MOUNTAINS · INTERVAL 200 m