How this site was made · written 9 Nov 2026, updated 10 Nov 2026
What ten prototypes taught us about the catalogue
Building ten working prototypes on the public catalogue and the live MCP server surfaced a handful of things worth knowing. Each one is traced to the page where you can see it.
These are the engineering notes. Who runs the site, what is live and what happens to the boards and trips you save is on About this site.
The one ask
Give listings coordinates, even rounded to 100 m
Listings carry only a neighbourhood name and a city. Every map, walking distance and "what is nearby" in the lab is built on geocoded neighbourhood centres, which can sit kilometres from the house. A latitude and longitude per listing would make the neighbourhood analyzer work for every listing, let search rank by distance, and remove a whole class of wrong-continent errors.
More under "Listings have no coordinates" below; see it on the neighbourhood analyzer.
About the MCP server
Host fees are invisible to every MCP call.
Coliving.com adds per-listing "other fees" on its own pages: on 10 Nov 2026 Pau Claris IV (Barcelona) showed "$1,011 /month + $449 other fees" on every room and Don Ramón De La Cruz II (Madrid) "$899 /month + $719 other fees", while the search cards, get_listing and check_availability returned no fee field at all for either. Whether the extra is per month or per stay is not stated on the listing either. A guest who was told "$893, right at your budget" cannot book at $893. Three of the five listings a reviewer checked that day carried no such fee, so it is per listing and nothing on the API side says which. Since 10 Nov every price on this site carries "plus any host fees shown on the listing" and no page says a place fits a budget without that line; a fee field on the card and in get_listing (amount, and whether it is per month or per stay) would let an assistant quote the real number.
Checked on coliving.com/spaces/fvlwyl0d and coliving.com/spaces/rjiek2cx on 10 Nov 2026; seen on the home page, friends and agents.
A plain search shows 1,241 of 2,067 published listings.
An assistant that does not pass availability_scope: "all_published" never sees 40% of the catalogue. On 9 November 2026 (02:53 UTC) the unfiltered search reported 1,241 listings in its default scope ("available within 30 days") and 2,067 with all_published; every default-scope listing is in the larger set. City counts follow the default scope too: search_locations reported 141 listings for New York City against 304 published under that name, and 56 for Paris against 188. The lab's own 27 October snapshot was the default scope as well, which is why it held 1,142 listings; every page now uses the full 2,067.
Measured by the lab's catalogue pull on 9 Nov 2026; live on agents.
The three defects found on 27 October are fixed.
Currency labels are now USD, country is filled on every card, and the unfiltered search no longer stops at 1,000. The agents prototype re-checks this every hour and shows the values it got back, with the 27 October value beside it.
Live on agents.
Budget filters return rooms above the budget.
"Shared room under $700" returned 11 of 12 cards above $700, because the filter looks at a different price than the card shows. The agent demo now counts over-budget results and says so.
Live on agents.
Two tools disagree on the same room's price.
For the same room on 9 Nov 2026, get_listing says $1,035 and check_availability says $1,089 (on 27 Oct: $1,052 against $1,208). Agents will quote one or the other.
Live on agents.
Deposit amounts have left the cards.
On 27 October 742 listing cards carried a deposit above zero; on 9 November none carries a dollar amount (505 say no deposit, 1,561 give no amount, one gives a percentage). Anything that adds a deposit to the cost of a first month now has to say it does not know it.
Measured by the lab's catalogue pull on 9 Nov 2026; affects the calculator.
About the data
The "from" price is often a dorm bed, even when the card says private rooms.
For houses whose room list has both, the headline price is the cheapest bed. In 448 of the 994 listings with a private-bathroom room, the cheapest such room costs more than the "from" price. A visitor who wants a private bathroom sees a number they can't book at. The where-next and calculator prototypes price the room the visitor actually asked for and say so under the card.
Seen in where next, calculator, friends.
Room-list prices and card prices come from different moments.
One Mexico City listing shows $1,499 in its room list, while its card said $1,942 on 27 October and $1,824 on 9 November. Any tool that mixes the two will make coliving for two look cheaper than for one, which is exactly what the first version of the calculator did. The fix is to scale room prices by their ratio to the listing's cheapest room, and to state it.
Seen in calculator.
Reported internet speed is uneven, and absent speed is not slow speed.
1,274 of 2,067 listings report a download speed; Boston reports it on 0 of 130, Taipei on 0 of 8. A relocation tool that treats a missing value as 0 Mbps buries whole cities. The relocation finder counts only reported speeds and says how many listings did not report one.
Seen in relocation finder.
Listings have no coordinates.
Only a neighbourhood name and a city. Geocoding those gave a pin for 1,516 of the 2,067 listings, 1,066 of them at neighbourhood level, but a neighbourhood centre can sit kilometres from the house, and anything built on top inherits that. The review pass found pins on the wrong continent for three places whose names exist elsewhere, which is what happens without coordinates at the source. A latitude and longitude per listing, even rounded to 100 m, would unlock maps, distances and nearby-life features across the board.
Seen in neighbourhood analyzer, year planner.
Community tags from hosts are the best signal for "who is this for", and they are under-used.
Host-set tags such as Surfers, Students, Founders and Nature lovers, together with the community-events list, place 1,416 of the 2,067 listings into readable activity collections with the reason shown on every card. Review summaries almost never mention hobbies; descriptions are not usable at all through the MCP.
Seen in hobbies.
The same place appears under several names.
Country is now filled on every card (39 listings had none on 27 October), but suburbs and spellings still split cities: Saint-Denis, Issy-les-Moulineaux and 14 more communes sit apart from Paris, Antwerpen apart from Antwerp. Bali does not exist as a searchable place; its 15 listings sit under ten names, from Kuta and Canggu to Kabupaten Badung and Kota Denpasar. The lab folds 38 such names into their city with one shared table, so every page counts the same cities.
Photos are served at upload size.
A grid of 30 cards weighed 11 MB from the CDN. The lab generates 480 px and 160 px WebP thumbnails once and every page uses them. A resized image variant from the CDN would do the same for every integrator.
Shared across all ten; first measured on wishlists.
If I had to pick three to take further
Where should I live next
It answers the question people actually type into a search engine, ranks on data you already have, and hands over a visitor with a budget, a month and a city. Add real climate normals and it is a landing page per city.
Coliving for your hobbies
Ten collections, each a page a surf forum or a founders' newsletter would link to. The classifier is rules you can read and fix. It needs host tags to be a little richer, which is a form field, not a model.
The agent side of your MCP
Assistants will search for places to live whether or not the data is ready. The hourly checks on the agents page are a monitor you could run yourself, and the two open traps are small fixes with a large effect on what an assistant tells its user.
Everything above comes from the public catalogue (2,067 listings, every published one, pulled with their detail text on 9 November 2026; the 27 October snapshot for the before-and-after numbers) and from live MCP calls on 8 and 9 November 2026, made by the lab's own code. No language model ran on the server; the explanations on each page are either rules shown next to the result or text written in advance and dated. Where a number is an estimate, the page that shows it says so.