Perplexity Computer Built a Map of 25,868 NYC Restaurants: Where the Data Came From and What It Shows

Perplexity's Computer account posted on October 3, 2026 that its agent had built a "3D map of nearly 26,000 restaurants and cafes across all five boroughs of New York City," where you can search by dish or neighborhood and then step inside places like Peter Luger and Grand Central Oyster Bar. A follow-up said it was built with OpenStreetMap, MapLibre GL JS and OpenFreeMap, and linked the live app.

Demo posts from agents are easy to wave at and hard to evaluate. This one is unusually checkable, because the app documents its own method. We opened it, read its Coverage and Credits pages, and compared them with the claim. The result is a useful case study in what agent-built apps look like when the data work is done properly.

TL;DR — what people are asking

What the app actually is

Window Seat opens on a dark map of the city where every point of light is a food-service permit. A side panel shows the count, "25,868 places lit across five boroughs," a row of "step inside" rooms, and a borough breakdown. You can search places, dishes and neighborhoods, save places, plan an outing, and compare venues.

The borough totals shown by default are Manhattan 10,163, Brooklyn 6,726, Queens 5,763, the Bronx 2,277 and Staten Island 939, which sum to 25,868. The "step inside" feature opens rooms for places including Katz's Delicatessen, Grand Central Oyster Bar, McSorley's, Caffe Reggio, Peter Luger Steak House and Gage and Tollner. The app labels them as built from real photographs.

On the "3D" description: the tweet calls it a 3D map. What we saw is a two-dimensional point map with photo-based, walk-in rooms. If the rooms are your idea of 3D, the claim is fair; if you expected extruded buildings, check the app yourself, because we did not verify any 3D rendering beyond the rooms.

The data funnel: from 295,568 rows to 25,868 places

The most impressive thing about Window Seat is its Coverage page, titled "How this atlas was built." It lays out a funnel from raw data to what is on screen:

This answers the question a reader asked under the tweet: how did it get the data? The answer is not a scrape of a review site. It is the city's public inspection dataset, cleaned and deduplicated, with every gap accounted for.

The page also says plainly what the map is not: "This is not 'every café in New York.' It is every place with a city food-service permit that [the health department] has on file, cleaned into one record per place, plus a small set of places researched by hand." Hidden by default are places not yet inspected, with no inspection in two or more years, closed at the last inspection, private or institutional, airport terminals, seasonal stalls, ticketed venues and delivery-only kitchens, and each group can be revealed.

So the tweet's "nearly 26,000" matches the default view of 25,868, and the app's own fuller count is 30,931 venue records.

Sources and credits

The Credits page is as thorough as the Coverage page. In summary:

• Map. Basemap tiles and fonts from OpenFreeMap using the OpenMapTiles schema; map data and the walking network from OpenStreetMap contributors under the Open Database License; rendering with MapLibre GL JS.
• City data. The health department's restaurant inspection results, outdoor-dining locations from the transportation department, building data from the city's PLUTO dataset, neighborhood and borough boundaries, and the city's GeoSearch for geocoding.
• Researched profiles. Ratings and distinctions from Tripadvisor, The Infatuation and the MICHELIN Guide, each linked and read-dated, with one-line summaries described as the atlas's paraphrases. Ratings stay with their sources.
• Photographs. 190 photographs across 73 researched places, each shown with its Wikimedia Commons license, such as CC BY or CC BY-SA, and a note where the room may have changed. Photos are marked as not endorsements by the photographers or venues.
• What is not used. Google Maps, Apple Maps and OpenStreetMap are linked for directions only, and no Google data, photos, reviews or ratings appear.

For anyone who has built a map app, this is the right shape: open data for the backbone, clearly licensed imagery, and linked rather than copied third-party ratings.

What it says about agent-built apps

Computer is Perplexity's agent product. We have covered its projects and multiplayer features, its orchestration costs, the GPT-5.6 Sol price cut and the portable Computer for local agents. This map is a good example of what such an agent can ship in a single project. Four lessons stand out.

1. The hard part was the data, not the map

MapLibre and OpenFreeMap make a basemap easy. The work was deduplicating nearly 300,000 inspection rows into one record per place, geocoding the missing ones, checking coordinates against borough boundaries, and flagging 12 records where the listed borough disagrees with the coordinates. An agent that does this well saves days. An agent that skips it ships a pretty map of bad data.

2. Publish your funnel

The Coverage page is the reason this app is credible. Every count is explained. If you build something similar, publish a table of how many records you started with, how many you dropped and why. It preempts the first comment, "is that really everything?"

3. Be explicit about what the agent did not make up

The app states that interiors are not generated or reconstructed and that unknowns are shown as unknown. For an agent-built product, that is a trust feature. Readers will otherwise assume the worst about generated content.

4. Licensing is part of the build

Each photo carries a license, a photographer and a date. For a product that mixes open data, community photos and third-party ratings, a credits page is not decoration; it is how you stay on the right side of terms.

What we could not verify

• How much the agent did. Perplexity says Computer built the app. We cannot see the prompts, the iterations or the human edits, so we do not know how much was autonomous. The app itself does not mention Perplexity or AI.
• Accuracy of individual records. Inspection-derived data can contain closed venues, wrong coordinates and stale names, as the app's own flags suggest. We spot-checked the headline numbers, not the places.
• The "3D" claim, as discussed above.
• Longevity. It is a hosted demo; availability and data freshness are not guaranteed.

Build your own: a practical recipe

If this makes you want to try a city-data map, here is a workable path.

1. Pick a public dataset with stable IDs. Inspection, permit and license datasets are ideal because each row has an identifier.
2. Ask the agent to profile the data first. Row counts, duplicate keys, missing coordinates and bad values, before any map code.
3. Define "one place" explicitly. The app merged permits sharing a name and address, and kept every permit ID.
4. Geocode only the gaps, and validate results against known boundaries.
5. Choose an open basemap such as OpenFreeMap with MapLibre, and keep attribution visible.
6. Add depth selectively. Research a small set by hand or with the agent, and cite every fact.
7. Write the method page. Counts, definitions, exclusions, licenses.
8. Test it on a phone. Dense point maps are heavy; measure frame rate and load time.

For the workflow around this kind of build, our guides to loop engineering with coding agents and what vibe coding is cover the habits, and Anthropic's Opus 5.5 showcase shows how other agents are being used for browser-based builds the same week.

What people are asking

Is this a Google Maps replacement?

No. It is a curated atlas built from city inspection records. It does not use Google data, and it is not a live reservation or review platform.

Does a high count mean every restaurant is open?

No. The default view hides several groups, such as places closed at last inspection, but data can lag reality. Treat it as an exploration tool, not a guarantee of a venue's current status.

Can I reuse the data?

The underlying city datasets are published under NYC Open Data terms and OpenStreetMap under the Open Database License. The app lists its sources and photo licenses; follow each one for your own use.

Is this proof agents can build whole products?

It is one data point. A single polished demo does not show success rates, cost, or how many tries it took.

Where is the code?

We did not find a public repository linked from the app or the posts.

Honest limitations

• We read the app's Coverage and Credits pages and compared them with the tweet; we did not audit individual venue records.
• Counts reflect the app on October 3, 2026 and may change.
• Perplexity's authorship claim comes from its own social posts; we could not confirm it independently.
• We did not capture screenshots of the app for this post because it contains third-party photographs.

Bottom line

Window Seat is a well-documented example of an agent-built data product: 25,868 places from city inspection records, 73 researched profiles, and 16 rooms built from licensed real photographs, with the method published on the page. The map stack is the easy part. The credit goes to the funnel, the sourcing and the honesty about coverage. If you build something similar, copy those habits.

Related on explainx.ai

• Perplexity Computer projects and multiplayer
• Perplexity Computer orchestration and cost
• Perplexity cuts Computer prices with GPT-5.6 Sol
• Perplexity Portable Computer for local agents
• Loop engineering with coding agents
• What is vibe coding?
• Opus 5.5 browser-build showcase

Source: the Window Seat app and its Coverage and Credits pages, as read on October 3, 2026, and the Perplexity Computer posts on X.

Counts and descriptions reflect the live app on October 3, 2026 and Perplexity's posts; both may change.