Mireye Launches a Physical-World Data Layer for AI Agents
Mireye, a Y Combinator S26 batch company, has launched what it calls the "physical-world layer for AI agents": a single API and MCP server (a protocol that lets AI agents call external tools) that returns sourced, enriched geographic data for any US coordinate. The product went live via a Launch HN post on Hacker News (source) and is documented at docs.mireye.ai, with a skills file at mireye.com/skills.md designed for direct paste into an agent's context (source).
The core problem Mireye addresses will be familiar to anyone building place-aware agents: today's large language models hallucinate when asked specific questions about specific places. A Fortune 500 insurer told Mireye's founder, Ansh, that its engineers had abandoned underwriting agents for exactly this reason. Ansh hit the same wall earlier while building construction agents that could reason about online content but had no knowledge of the physical ground under a location (source).
Mireye's response is an infrastructure play rather than an application. The company's first product was a niche site-screening app, which Ansh discontinued after discovering that 311 of 317 catalog fields were queried with no dominant use case. The signal was clear: customers wanted the underlying data layer, not a vertical tool (source).
That layer now comprises 366 fields served multi-tenant from a single index, drawing on 85 authoritative sources. The catalog is mostly US federal data, sourced from agencies including USGS, FEMA, NOAA, USDA, BLM, EPA, and EIA, with additions such as Overture Maps and Sentinel-2 imagery. Coverage spans the full United States, with drive-time analysis (via /v1/proximity) additionally covering Canada (source).
What separates Mireye from a raw data dump is its approach to provenance and what the team calls "absence-typing." Every field in every API or MCP response carries four metadata elements: source name, source URL, a fetched_at timestamp, and a confidence level. Each field also returns one of three statuses: ok, absent, or failed. The distinction matters: an absent status tells an agent that a flood zone genuinely does not exist at a coordinate, while failed indicates that the county never mapped floods at all. For an agent making a downstream decision, that semantic gap is consequential (source).
Data freshness varies by layer. USGS 3DEP elevation data is effectively static. FEMA flood maps refresh on FEMA's own cadence. USDA cropland data updates yearly. Mireye gathers each source county by county in whatever format a given county publishes, normalizes everything into one schema, and contracts and refreshes continuously. Where a needed dataset does not exist in any index, Mireye has filed public records requests to obtain it, as it did for sewer main data in North Carolina (source).
The company also offers on-demand indexing. A long-running agent researches candidate sources, collects data, tests it against ground truth, and indexes the requested field, usually within a day. Every plan includes field requests, allowing users to order fields that do not yet exist. Mireye researches, builds, and validates the field, then notifies the user when it goes live permanently (source).
Pricing is credit-based with a free tier of 5,000 credits per month requiring no card. Paid plans start at $19 per month. No plan bills past its allowance; usage stops rather than accruing overage charges (source).
Mireye's stated mission is to "index every inch of the earth and make it as queryable as the web" (source). The current catalog covers the US, with limited Canadian coverage for proximity analysis. The company is listed under both Y Combinator's Artificial Intelligence and Infrastructure startup categories (source).
In my view, the design choices here reflect a genuine understanding of where agent failure modes actually live. The combination of per-field provenance and explicit absence-typing addresses a specific, documented problem: agents that confabulate place-specific facts because their training data contains no ground truth for a given coordinate. The on-demand indexing model is the more ambitious bet. If it scales beyond sewer mains in North Carolina, it turns Mireye from a static catalog into something closer to a physical-world crawler, filling gaps in the long tail of county-level data that no single provider has incentives to normalize. Whether that model holds at volume is an open question, but the early customer list spans insurance portfolio screening, warehouse sourcing, data center site selection, drone deployment planning, and signals for human trafficking investigations, which suggests the demand for queryable ground truth is broad and not confined to a single vertical.
