Three kinds of product sit on the same H3 cell, so any two pieces join without a spatial operation. The data itself, the knowledge bases that document how the grid works, and the MCP services that let an agent do the conversions.
Tables and vectors in your warehouse. Every card returns a Manhattan sample.
250 Census attributes per cell. The free context floor, on BigQuery Marketplace, no signup. Population, income, education, housing and commuting, already projected onto the grid.
The full hourly context for any US metro. Every cell, every hour of the week. It tells you who is around a place on a Tuesday at 10am rather than on average. Manhattan is free.
The vector embedding of real-world context. A learned representation your model queries directly, instead of a set of hand-picked columns.
Brings your own data into our latent space and aligns it with 168 VE, so your vectors and ours speak one language.
Where the people in a cell come from, weighted by commute flow and sector, built on LEHD origin destination records.
Employment by NAICS sector for each hour of the week, so a retail cell and a hospital cell stop looking alike at 9pm.
The probability that a business of a given type is open in a cell, by hour class, without scraping storefront pages.
A label that groups cells behaving alike, so a plan can run on a few dozen place types instead of millions of individual cells.
How many people are in a cell in each hour of the week. Resident and worker counts on a UTC axis, so Tuesday 10am in Denver lines up with Tuesday 10am in Boston.
Which TV and FM stations cover a cell, counted from FCC service contours. It measures coverage capacity rather than viewership.
Transit stops per cell by mode, from the National Transit Map. It counts stop presence rather than boardings.
Daily weather flags per cell, scored against local normals rather than fixed thresholds, so a hot day in Phoenix and a hot day in Boston mean the same thing.
When resident workers leave a cell in the morning, in the ACS departure bands. The bands are local clock, so pair it with the Time Crosswalk before joining an hourly table.
School and university sessions by state and campus, with SNAP issuance days. Session windows follow state norms rather than district calendars.
Boundary to H3 and back for county, tract and block group, with the overlap fraction on every edge so a rollup is auditable. Media market rollups are not included, because those boundaries are licensed from the rating provider.
Any timestamp to its hour of the week, correct across time zones and daylight saving, which is where most hourly joins quietly go wrong.
Where a screen could sit: 2,174,623 venues across 16 classes, counted per cell. It measures venue supply rather than screens sold or seen. Overture leaves operating status unset on 686,458 of them, and those are retained and counted rather than dropped.
Where purchase occasions physically happen: 2,168,244 retail outlets across 24 classes. It records store presence rather than sales or media participation. Brand identity is deliberately absent, so a chain is a count rather than a name.
261,702 lodging properties across 8 classes, counted per cell. Property counts only, with no room counts, so it sizes lodging supply rather than bed capacity.
The written standard behind the grid. Open, no signup.
The written standard for moving between boundaries and the H3 grid. It covers inscribed and circumscribed circles, containment rules and weighted crosswalks, with the conversion code tested against every claim on the page.
The same treatment for time. It defines one hour of week axis and catalogues 62 edge cases across 13 categories where an hourly join silently produces the wrong answer.
MCP servers. Your agent connects once and calls the tools directly.
The knowledge base as a tool surface. An agent connects once and runs the tested conversions live, rather than reading a page and reimplementing it. It covers both corpora, geographic and temporal.
The planning layer as tools, run against one client's store network. Each tool returns the audience around the stores, the people in reach of them, on the H3 grid. The experiment tools design matched test and control markets and read the lift against the client's own point of sale.
The knowledge bases need no request. They are open links. For the data and the services, a request gets you this. Nothing to install and nothing to move, since a sample is a table in your warehouse, or a file if you would rather look at it first.
You pick a card and send the request. The subject line carries the layer name, so it routes without a call first.
A Manhattan slice of that layer with the schema, the column definitions, and one query that answers a question you care about.
If the sample holds up, we scope the metros or the states you need and price from there.
