Hacker News 熱門精选77

发布 HN:Mireye (YC S26) – 物理世界 AI Agent 基础设施

Launch HN: Mireye (YC S26) – Infrastructure for Physical World AI Agents

· 40天前
跳到正文

精选理由

帮读者了解物理世界AI代理的地点决策基础设施

AI 摘要

大家好,HN,我是 Ansh,Mireye 的创始人。我正在构建 AI agents 用来对物理地点做决策的基础设施:针对任何美国地点的数据、增强、工具和信号,全部通过一个 API 和 MCP server 提供。 这里有一段演示视频: https://youtu.be/haqO6UbUqU0 要试用的话,把 https://www.mireye.com/skills.md 粘贴到你的 agent 里,然后到 mireye.com 领取一个免费 key(5,000 credits,无需信用卡)。文档在 https://docs.mireye.ai 。最快的上手方式是让 API 回答任何关于美国某个地点的问题。找一个你熟悉的地方来测试,把结果冷静地和你已知的信息对比打分。 在 Mireye 之前,我自己在做建筑 agents 的过程中也撞上了这堵墙:我的 agent 也许能推理任何线上信息,但对其脚下的地面一无所知。后来一家 Fortune 500 保险公司告诉我,他们的工程师也因为同样的原因放弃了核保 agents。前沿模型在被问到某个特定地点的具体问题时,仍然会胡编乱造。 我的第一个产品是一个小众的场地筛选应用。客户拿他们熟悉的地点来测试,答案表现不错,但没人关心这个应用本身。他们想要的是底层引擎。使用情况也印证了这一点:目录中 317 个字段里有 311 个被查询,而且没有哪一种用例占主导。所以我砍掉了应用,转而开始构建基础设施。 Mireye 不是一个上面套了 API 的数据集,因为事实本身并不等于决策。agent 通过它完成整个工作:带来源标注的事实、把一个普通地址增强成所有者、面积、建筑物和附近电力设施的地址信息、让运营模型容易出错的工具,以及当某些事情发生变化时的信号,比如重新分区申请。数据、增强、工具、信号。 这些工具来自对 agent 失败的观察。我们基于自己的基础设施构建 agent,而同样的事情不断出错:agent 会凭目测估算距离,而不是计算;会把地址对应到错误的地块;或者在批处理到一半时把全部预算耗光。每一种失败都变成了 agent 可以调用的东西:确定性的几何和行车时间工具、地块解析、在任务运行前给出价格的 quote endpoint,以及把完整工作流打包起来的 skills,比如场地筛选或地址核保。 最难的部分出乎我的意料。每个数据源都必须被采集进来(有时要按县逐个处理,而且每个县发布的格式都不同)、归一化到同一 schema、注明约束条件(它来自哪里、每个值是什么意思、多久刷新一次),然后永远保持最新。我们今天为 366 个字段运行这个循环,并通过一个索引以多租户方式提供服务,而且每个数据源的抵抗方式都不同。Maryland 发布了一个数据集,标题就叫“Hidden Property Owner Names”。我在 North Carolina 提交了公开记录请求,因为没人会索引下水管主管道。 更深层的问题是语义。两个县发布一个同名字段,但含义却不同。而最危险的值是 null。它的意思是“这里没有洪水区”,还是“这个县从来没绘制过洪水区”?把模型放在这个空白前,它就会用一个看似合理的数值把沉默填满。所以我们要为缺失建模。每个字段都会返回 ok、absent 或 failed。客户告诉我,正是这种拒绝回答的机制让他们信任它。 最新的部分是按需索引。如果你要一个我们还没有的字段,一个长期运行的 agent 会去研究来源、收集数据、用真实地面情况进行测试,然后把它编入索引,通常一天内就能完成。我会从这个帖子里挑几个字段请求,并回来汇报它构建了什么。 有人已经基于它做出了我没预料到的东西:保险团队用它筛查洪水、风灾和野火风险的组合资产;一家 proptech 通过增强清洗杂乱的房源地址;一个健康品牌用它给街角的海报点位打分;一家机器人公司用它寻找仓库;数据中心选址;无人机部署规划;某城市的校车路线;以及用于人口贩卖调查的信号。我们的使命是把地球上的每一寸土地都编入索引,并让它像 web 一样可查询。 定价是公开的:免费套餐每月 5,000 credits,$19/月可得 25,000,$99/月可得 120,000,企业版定制。 我很想听听你们正在构建的 agent 会在现实世界的哪些地方落地。如果你能用 Mireye 在一个你熟悉的地方跑一下,并告诉我哪里错了,我会更开心。 Comments URL: https://news.ycombinator.com/item?id=49552616 Points: 2 # Comments: 0

正文

Physical world context for AI agents

One API for data, enrichment, tools and signals.

Sign upBook a Demo

Every value cited

  • USGS
  • FEMA
  • NOAA
  • USDA
  • EPA
  • EIA
  • FCC
  • CENSUS
  • NREL
  • USFWS
  • USFS
  • FHWA
  • FAA
  • BTS
  • USACE
  • BLM
  • HUD
  • FHFA
  • BLS
  • NSIDC
  • OVERTURE
  • REGRID
  • JRC
  • SENTINEL-2
  • USGS
  • FEMA
  • NOAA
  • USDA
  • EPA
  • EIA
  • FCC
  • CENSUS
  • NREL
  • USFWS
  • USFS
  • FHWA
  • FAA
  • BTS
  • USACE
  • BLM
  • HUD
  • FHFA
  • BLS
  • NSIDC
  • OVERTURE
  • REGRID
  • JRC
  • SENTINEL-2

Power your agents to make decisions about the physical world

Without Mireye

What's the elevation at 40.7128, -74.006?

Claude

I don't have real-time access to elevation databases. Based on general knowledge, the area near 40.7128, -74.006 (Manhattan, NYC) has elevations ranging from roughly sea level to about 10 to 50 meters depending on the specific location.Would you like me to suggest some online tools for looking up precise elevation?

No source. No timestamp. Estimated range.

With Mireyelive · captured 2026-07-28

What's the elevation at 40.7128, -74.006?

Claude, powered by Mireye

The elevation at 40.7128, -74.006 is 13.03 meters above sea level (NAVD88 datum).

elevation
= 13.03 meters
fetched
= 2026-07-28
confidence
= medium

Federal source, clickable URL, fetched timestamp, confidence.

One API to connect your agents to physical world data, enrichment, tools and signals

AI agents — Claude, ChatGPT, Kimi, Gemini, Cursor, Grok, Copilot, DeepSeek, or custom agents — ask Mireye through one API with MCP tools. Mireye returns cited data, enrichment, proprietary fields, signals and tools, indexed from the physical world. A field the catalog doesn't have yet can be requested — matched against the catalog first, and a genuine gap becomes one queued build that flows back into Mireye for every caller after.

AI agents

asks
Mireye

One API · MCP tools

What you get

  • Cited data
  • Enrichment
  • Proprietary fields
  • Signals
  • Tools
indexed

Physical world

Field requests, queued builds
source data

AI agents

asks
Mireye

One API · MCP tools

What you get

  • Cited data
  • Enrichment
  • Proprietary fields
  • Signals
  • Tools
indexed

Physical world

source data
Field requests, queued builds

→ back into Mireye

Sectors we power

Every “should I care about this place, and why?” question routes to the same API — these are the desks already asking it.

  • Data center siting

    Ashburn, VA

    Screen a data center at 39.0438, -77.4874: slope, FEMA flood zone, nearest 345 kV line, substation voltage and status, and the county's interconnection queue.

    Mireye

    • Power
    • Flood
    • Terrain
    • Queue
    Show the agent question for Data center siting
  • Renewable energy siting

    Central Iowa

    Solar farm screening at 41.9, -93.5: slope, cardinal aspect, annual irradiance, interconnect distance, and the soil drainage class.

    Mireye

    • Slope
    • Irradiance
    • Interconnect
    • Land cover
    Show the agent question for Renewable energy siting
  • Insurance underwriting

    Galveston, TX

    Coastal exposure at 29.2899, -94.7875: flood zone, base flood elevation, design wind speed, distance to coast, wetland intersect.

    Mireye

    • Flood
    • Wind
    • Coast
    • Wetlands
    Show the agent question for Insurance underwriting
  • Mortgage & title

    Lower Manhattan, NYC

    Property hazard disclosure at 40.7128, -74.006: elevation, coastal proximity, watershed jurisdiction, wetland intersection.

    Mireye

    • Terrain
    • Coast
    • Watershed
    • Wetlands
    Show the agent question for Mortgage & title
  • Residential land

    Williamson County, TX

    Diligence a 140-acre tract at 30.6327, -97.6772: parcel boundary and acreage, floodplain overlap, distance to transmission, wetland acres on the parcel.

    Mireye

    • Parcels
    • Floodplain
    • Transmission
    • Wetlands
    Show the agent question for Residential land
  • Commercial lending

    Phoenix, AZ

    Collateral assessment at 33.4484, -112.074: slope, flood zone, road access, utility service territory, soil shrink-swell class.

    Mireye

    • Flood
    • Access
    • Utilities
    • Soils
    Show the agent question for Commercial lending

Ask anything about a place.

Ask a question about a place and get a cited answer. One call can resolve an address, fetch fields, and measure distances, and you get back the plan it ran so you can replay it.

mireye powered agent · claude-opus

▍

Turn anything into a place.

Send a US street address and get back the coordinate it names, with how that coordinate was derived — sitting on the parcel, or interpolated along a street centerline — and a confidence score. Centroid-grade and low-confidence matches are refused rather than guessed.

mireye powered agent · gpt-5 · tools: mireye_geocode

▍

One string, the whole stack.

Geocoding gives you a point. Lookup gives you the place: county, tract, congressional district, metro, timezone, elevation and flood status resolved in the same call — and a typed refusal when the match is not good enough to stand behind.

mireye powered agent · gpt-5 · tools: mireye_lookup

▍

Cited facts at any location.

Name the fields or a preset and get typed values for any US location, each with its source, fetch time, and confidence. Send hundreds of locations in one call, and one bad address does not fail the rest.

presetslong jobs — /v1/runs
  • boundaries · 4
  • building_lookup · 4
  • data_center_siting · 90
  • flood_risk · 13
  • grid_interconnect · 29
  • land_cover · 5
  • natural_hazard · 14
  • points_of_interest · 23
  • site_selection · 54
  • solar_siting · 26
  • storage_siting · 19
  • terrain · 6
  • utilities · 18
  • wildfire_underwrite · 6
  • wind_siting · 26

mireye powered agent · claude-opus · tools: mireye_fetch

▍

Measure what the data can't say.

Four drive-time operations at any US or Canadian location: point-to-point distance, nearest-by-road over curated infrastructure sets — airports, substations, power plants, rail, ports, urban areas — screening by drive-time cutoff, and labor sheds. Every answer states what it charged and what it can't know.

mireye powered agent · gpt-5 · tools: mireye_proximity

▍

Request a field that doesn't exist yet.

Describe the data in plain language, with the locations where you need it. Already in the catalog and you get the field plus a cited sample now; close but not exact and you get the near miss to accept or reject; nothing answers it and one build is queued with a request_id to poll. An ask Mireye can't index — real-time, commercially licensed, personal contact data — comes back as a typed no with somewhere else to go.

mireye powered agent · gpt-5 · tools: mireye_request_field, mireye_field_request_status

▍

Frequently asked questions

What does Mireye actually do?

Mireye is the physical-world layer for AI agents: sourced data, enrichment, and the tools on top of them. Ask a question in natural language, resolve an address to its canonical parcel, or fetch cited fields at any US coordinate — one API and one MCP server, with a citation attached to every field.

What does it cost?

Credit-based pricing. The free tier is 5,000 credits a month with no card; paid plans start at $19/mo, and no plan bills past its allowance — usage stops rather than running up an overage. Credits are the unit across all endpoints, and the credit price per call type is published on the pricing page.

What region does Mireye cover?

United States for the data catalog, with one exception: drive-time analysis (/v1/proximity) covers the US and Canada. Most of the catalog is US federal (USGS, FEMA, NOAA, USDA, BLM, EPA, EIA and more), with a few open and licensed additions such as Overture and Sentinel-2. Every value carries its own source, so you can always see which one you are relying on.

Where does the data come from?

85 authoritative sources, federal datasets primarily, plus county records and open datasets. Every field in every response returns a source name, a source URL, a fetched_at timestamp, and a confidence level so the agent can pass provenance straight through to the end user or auditor.

How fresh is the data?

It varies by layer. Elevation (USGS 3DEP) is effectively static. FEMA flood maps refresh on FEMA's cadence. USDA cropland updates yearly. Every response includes the fetched_at timestamp so the agent always knows what it is looking at.

What if the field I need doesn't exist?

Order it. Every plan includes field requests — file one and we research, build and validate the field, then notify you when it goes live: permanently, for every future caller. Requests are matched against the catalog first, and honestly rejected when they can't be built.

Give your agents the context to act in the physical world.

PricingBook a Demo
查看原文(Hacker News 熱門)