Skip to main content
An AI agent knows the internet. It does not know the ground. Ask a frontier model how much power it can get at a specific parcel, what the county will let you build there, or whether the wetland on the north edge kills the site, and it will estimate. Mireye answers — from the authoritative record, with the source, the URL, the fetch timestamp, and a confidence bucket attached to every value. That is the whole company: the infrastructure layer agents call when the question is about a real place.

​
Four things behind one API

Mireye is not a map, a GIS tool, or a data reseller. Four distinct capabilities ship behind the same auth, the same catalog, and the same provenance model.

Data

Cited facts at a location. Public agencies (USGS, NOAA, EPA, EIA), transparent open data (OpenStreetMap/OpenInfraMap, Overture), and explicitly named licensed sources — normalized to one schema. POST /v1/fetch

Enrichment

Derived answers no single source publishes: deliverable power bands, developable acres net of wetland and easement, parcel resolution from a messy string. POST /v1/lookup

Tools

Computation over locations, not lookups at one: drive-time matrices, nearest-candidate search, proximity screens, labor sheds. POST /v1/proximity

Coverage on demand

The catalog is not fixed. Ask for a field that doesn’t exist and a genuine gap becomes a queued build that lands for every caller after you. POST /v1/field-requests
Agents reach all four through natural language with POST /v1/ask — a planner selects the fields, fetches them in parallel, and a synthesizer writes the answer with inline citations — or through the MCP server, which exposes the same operations as native tools to Claude, Cursor, and custom agents.

​
The catalog is mostly infrastructure

These counts track the catalog build in this repository. The deployed API is the authoritative list — read it from GET /v1/meta/fields, which is public and unauthenticated.
  • Fields: 366 across 7 layers (terrain, land cover, built environment, utilities, parcels, climate, hazards). Full catalog at GET /v1/meta/fields.
  • Presets: 15 use-case bundles (terrain, flood_risk, wildfire_underwrite, land_cover, site_selection, building_lookup, points_of_interest, utilities, boundaries, solar_siting, wind_siting, storage_siting, data_center_siting, grid_interconnect, natural_hazard).
The largest layer is not maps. It is utilities and grid — substations, transmission voltage class, interconnection context, water service areas, wastewater plant capacity, fiber. That is the layer that decides whether something can actually get built, and it is the layer nobody had normalized. Check the split yourself — the catalog endpoint is public, no token:

​
What we won’t do

No black-box scores

No proprietary risk score, no “AI-derived” number dressed as data. Every value traces to a record you can re-fetch and verify.

Absence is an answer

Where the record doesn’t exist, you get absent with the reason — not a plausible guess. Source failures surface in partial_failures with a retryable flag.
An agent that gets a confident wrong number about a place makes a worse decision than an agent that gets told the data isn’t there. We optimize for the second.

​
Coverage

/v1/ask, /v1/fetch, and their MCP tools are United States only. Primary bounds are lat ∈ [18, 72], lng ∈ [-180, -65]; the Western Aleutians additionally accept lat ∈ [51, 54], lng ∈ [172, 180). Out-of-bounds requests return 400 coord_out_of_bounds. Drive-time tools in /v1/proximity additionally cover Canada; other endpoint pages document their own locator coverage.

​
Start here

Quickstart

First cited answer in five minutes. curl and jq, no SDK.

MCP install

One command to give Claude Code or Cursor the physical world.

Pricing and credits

What each call costs, and how to quote a job before you run it.

API reference

Every endpoint, with try-it panels.