Mireye suits agent builders who need real-world site context, such as flood zone, grid access or parcel data, without wiring up federal agencies one by one. It ships 15 presets and 7 MCP tools. It is US-focused and credit-metered, so it fits screening and diligence work better than global coverage or map display.
What changed
- · Release Mireye launches as YC S26: MCP server for physical-world AI agents (v1.0) source
Mireye serves cited facts about any US location to AI agents over a hosted API and an MCP server. The catalog holds hundreds of typed fields across 7 layers from 96 sources, each value carrying a source URL, fetch timestamp and confidence. Claude Code, Claude Desktop and Cursor connect through MCP, and missing data returns absent.
Maker: Mireye · Protocol: MCP · Auth: api key
Compatible agents: Claude Code, Claude Desktop, Cursor, Any MCP-compatible client
Required runtime: Mireye API key (free signup, no card), uv with uvx for the local stdio adapter (the hosted HTTP MCP endpoint needs no local runtime)
About Mireye
Mireye is a hosted API and MCP server that gives AI agents cited facts about a place in the United States. It was founded in 2026 by Ansh Chokshi and Shashwat Kapoor, who are part of Y Combinator's Summer 2026 batch in San Francisco. At its June launch the product covered five layers of federal data; it now serves a catalog of hundreds of typed fields across 7 layers (terrain, land cover, built environment, utilities, parcels, climate and hazards) drawn from 96 sources, and the largest layer is utilities and grid data such as substations and transmission voltage class.
An agent reaches Mireye through six endpoints: ask for natural-language questions, geocode and lookup for turning an address into a place, fetch for named fields or one of 15 presets (data_center_siting alone bundles 135 fields), proximity for drive-time calculations, and field-requests for data the catalog does not yet hold. Every value returns its source name, source URL, fetch timestamp and a confidence level. Where a record does not exist the API answers absent with a reason, and low-confidence address matches are refused instead of guessed.
The same operations are exposed as 7 MCP tools. Claude Code connects to a hosted OAuth endpoint with one command, while Claude Desktop and Cursor use a local stdio adapter that runs through uvx. Mireye sits beside other data skills an agent might call in the same session; web retrieval through Olostep is handy when an agent must read a county notice about the same parcel. Findings can be written to a spreadsheet with Google Sheets. Models from Anthropic and other labs can all call it, since the contract is MCP or plain HTTP.
The other side of the argument matters before you adopt it. Coverage is the United States, with only the drive-time tools reaching Canada, so an agent working in Europe or Asia gets nothing. The federal datasets underneath are free at source, so a team that needs one layer, such as elevation, can call USGS directly and skip the credits. Mireye also declines requests it cannot index responsibly, including real-time feeds, commercially licensed data and personal contact data, and parcel owner lookups are the most expensive call on the price list. It is a young product from a startup that launched in 2026, so support terms and uptime history are short.
Key Features
- A citation on every value: Each returned field carries a source name, source URL, fetch timestamp and confidence level, drawn from 96 sources.
- Catalog of hundreds of typed fields: Fields span 7 layers, and utilities and grid data (substations, voltage class, interconnection context) is the largest.
- 15 ready-made presets: Bundles such as flood_risk (13 fields) and wildfire_underwrite (10) each answer a whole screening job in one fetch.
- Seven MCP tools plus REST: Ask, fetch, geocode, lookup, proximity and field-request operations are exposed through a hosted OAuth endpoint and a local uvx adapter.
- Refusals instead of guesses: Low-confidence address matches are refused, and missing records return absent with a reason and a retryable flag on failures.
- Field requests: A field the catalog lacks can be ordered; a genuine gap becomes one queued build that serves every later caller.
Use Cases
- Data center site screening: An agent runs the data_center_siting preset on a candidate coordinate and returns slope, FEMA flood zone, nearest transmission line, substation voltage and interconnection context, each with its source.
- Coastal insurance underwriting: An underwriting agent pulls flood zone, base flood elevation, design wind speed, distance to coast and wetland intersection for an address and cites each figure in the file.
- Land tract diligence: A land-buying agent checks parcel boundary and acreage, floodplain overlap, distance to transmission and wetland acres on a tract before an analyst reviews it.
Install
claude mcp add --transport http --scope user mireye-earth https://api.mireye.com/mcpRequirements
- Sign up at mireye.com for an API key; the free tier needs no card
- Claude Code: authorise the hosted endpoint through the browser login that /mcp opens
- Claude Desktop or Cursor: install uv, then run mireye-mcp login or set MIREYE_BEARER_TOKEN
Actions
Ask
Answers a natural-language question about a US place with cited fields, and returns the plan it ran so the call can be replayed. MCP tool: mireye_ask.
curl -X POST https://api.mireye.com/v1/ask \
-H "Authorization: Bearer $MIREYE_API_TOKEN" \
-H "Content-Type: application/json" \
-d '{"question":"What would stop me building a data center here?","address":"2201 Seawall Blvd, Galveston, TX"}'question(string) — required: The question to answer about the place.address(string): Any US address. Give this or lat and lng.lat(number): Latitude in decimal degrees, used with lng instead of an address.lng(number): Longitude in decimal degrees, used with lat instead of an address.
Fetch Fields
Returns typed, cited values for named fields or a preset at one or many US locations, with one bad address not failing the rest. MCP tool: mireye_fetch.
lat(number) — required: Latitude in decimal degrees.lng(number) — required: Longitude in decimal degrees.fields(array): Field names to return. Use this or a preset.preset(string): A bundle such as flood_risk, terrain or data_center_siting.
Geocode
Turns a US street address into a coordinate and states how the match was derived, refusing coarse or low-confidence matches. MCP tool: mireye_geocode.
address(string) — required: A US street address.
Lookup Place
Resolves an address or coordinate to the surrounding place in one call, with a typed refusal when the match is not good enough. MCP tool: mireye_lookup.
address(string): A US street address.
Proximity
Runs drive-time operations at US or Canadian locations: point-to-point distance, nearest-by-road over infrastructure sets, screening by cutoff, and labor sheds. MCP tool: mireye_proximity.
Request Field
Describes data the catalog lacks in plain language and returns the existing field, a near miss, a queued build with a request_id, or a typed no. MCP tool: mireye_request_field.
Field Request Status
Checks the progress of a queued field build. MCP tool: mireye_field_request_status.
request_id(string) — required: The identifier returned by Request Field.
How to Invoke
Exposed as 7 MCP tools (mireye_ask, mireye_fetch, mireye_geocode, mireye_lookup, mireye_proximity, mireye_request_field, mireye_field_request_status) the agent calls by name, or as REST endpoints under api.mireye.com/v1 with a Bearer token.
Pricing
Plans read on the vendor pricing page on 2026-10-18, billed as monthly credit allowances: Free is $0 with 5,000 credits and 20 requests a minute; Build is $19 a month with 25,000 credits and 60 requests a minute; Growth is $99 a month with 120,000 credits, 300 requests a minute and a shared Slack channel; Enterprise is quoted with custom SLAs. No plan bills overage, since calls stop at the allowance, and unused credits do not roll over. Published credit costs per call: geocode 1, each field 1 per location, ask 10, proximity 12 per driving calculation, and parcel lookup 300, charged only on a verified match.
Strengths
- Provenance is built into the response, so a diligence file can cite the exact federal record and fetch date instead of a model's estimate.
- No plan bills overage, which caps spend at the plan price and makes agent loops safe to leave running on the free or Build tiers.
- Presets compress multi-agency screening jobs into one call, saving the days a hand-built pipeline over USGS, FEMA and utility data would take.
- The field-request route means a missing layer can be added to the catalog rather than worked around.
Weaknesses
- Nothing outside North America is covered, and even there only the drive-time tools reach Canada.
- Parcel lookup is the priciest call, so the free tier resolves roughly 16 parcels a month.
- Unused credits expire each period and requests stop at the allowance, which can interrupt a heavy month.
- It is a Summer 2026 startup, so uptime history and support depth are still short.
Frequently Asked Questions
How much do you pay for Mireye?
Mireye sells credits, and a plan is simply a monthly credit allowance. Build costs $19 a month for 25,000 credits, Growth costs $99 for 120,000, and Enterprise is quoted per account. As a yardstick, the vendor puts a full 135-field data center screen of one site at 135 credits, about 14 cents. Reaching the cap stops requests rather than adding a charge, so the plan price is the most a month can cost.
Can you use Mireye without paying?
Yes, within limits. A free account gets 5,000 credits a month with no card, a 20-requests-a-minute rate limit and one field request at signup. Credits reset on the 1st of each month (UTC) and do not carry over. Because a parcel lookup costs 300 credits, a free account can resolve only around 16 parcels a month.
Which tools compete with Mireye in 2026?
The nearest substitute is querying the agencies directly: USGS elevation, FEMA flood maps and the National Land Cover Database are free at source, though each has its own format and limits. Esri's ArcGIS APIs suit teams that want a full GIS platform with editing and maps, and Mapbox suits routing and map display. Mireye is the pick when an agent needs typed, cited values from many sources in a single call.
Is Mireye better than calling federal APIs directly?
For an agent that needs several layers at once, usually yes: one fetch returns many possible fields with source, timestamp and confidence, and gaps come back as an explicit absent with a reason. For one dataset on its own, such as elevation, the federal service is free and Mireye adds a vendor dependency for little gain. It is also a young product from a Summer 2026 startup, so check support and uptime terms before building a production path on it.
How long does it take to get going with Mireye?
The vendor's quickstart promises a first cited answer in about five minutes using curl and jq. Sign up on mireye.com for an API key, then call the REST endpoints with a Bearer token or add the MCP server. In Claude Code the hosted endpoint is registered with a single claude mcp add command and authorised through the browser login that /mcp opens. Claude Desktop and Cursor run the local uvx adapter and need mireye-mcp login or a bearer token.