From a chat message to an inbox of checked apartments

Fred is a tool-using AI agent with a web app around it. You chat with it in plain words, and it saves your search, starts crawlers on the listing sites, checks every building with the city and writes up the places that fit, while the app shows each step as it happens.

Why we built it. James and I built Fred at the AI Engineer NYC hackathon on October 11, 2026, because we’re hunting for an apartment ourselves and were tired of checking every site and alert email by hand. It had to be live by the end of the day, so we picked tools that ship fast, run without servers to look after, and let two people and their agents work at the same time without stepping on each other.

Role

I built the web app; James built the API, the agent and the crawlers

Stack

TypeScript · Next.js · Mastra · Claude · Hono · Zod · Cloudflare

Surface

Web app, chat and email alerts

Focus

Showing the agent’s work · nothing sent without you

• No listing site has an API

  StreetEasy, Zillow and the rest only show their listings on their own pages, so every source needs its own crawler.

• Search results leave things out

  A listing card rarely says the fees, the pet policy or the laundry, which are the things renters ask about first.

• The costly cases are the quiet ones

  A broker fee, a building with open violations or a scam deposit is exactly where a confident wrong answer hurts most.

• Trust needs receipts

  To believe an agent, you need to see what it checked, where each number came from and what it doesn’t know.

Four stages, with you in charge of anything that leaves the app.

• 1. You say what you want

  Fred turns “1 bed under $4,500 in Greenpoint, dog OK” into a saved search: dealbreakers, nice-to-haves with a weight, and notes in its memory.

• 2. Fred checks the sites

  It starts crawlers on StreetEasy, Zillow, RentHop, Zumper, Compass, Nooklyn, Craigslist and landlords’ own sites for that search. The same apartment on several sites becomes one place.

• 3. It checks each place

  Code drops places that clearly miss a dealbreaker. Fred reads the full listing page when facts are missing, looks at the photos, and scores the building from NYC Open Data.

• 4. It writes up the ones that fit

  Fred adds them to your inbox with why it picked each one, emails strong fits if you asked, and drafts messages to landlords that only send after you approve.

The loop

Fred is a Mastra agent on Claude Sonnet 5.5 at low effort, for speed. It works in a loop, calling tools and reading what they return, and it only writes its reply once the tools are done, so the places it talks about are ones it has actually looked up. The model is one setting, so the same agent runs on Gemini too, and 15 scripted conversations check what Fred did, not just what it said.

The hands

Every action is one of 25 typed tools with a Zod schema. The model never touches the database. It calls tools that read and write the same data the app shows, so whatever Fred changes, you see.

The wiring, and why

Everything runs on Cloudflare, so there was no server to set up or look after, and the app and the API share one domain.

Two people and a few agents

We each worked with a few Claude Code agents. They coordinated on a shared message board, where they posted contract changes, bugs and questions for each other, and the API contract lived in one place (packages/api) that both sides built from. CI checked the Worker, the app and the contract on every push, so a change on one side that broke the other failed the checks.

My side was the web app, and the hard part was making an agent’s work readable while it streams in.

• Live steps from the stream

  Each tool call arrives as its own part. I turn it into plain words (“Searching Greenpoint”, then “Searched Greenpoint”), refresh the inbox when a tool that changes it finishes, and fold the steps into one line. Next’s compression was holding the whole stream until the end, so every step showed up at once; turning it off in Next, since Cloudflare compresses at the edge, fixed it.

• A contract both sides build from

  The app’s types are generated from the Zod schemas, and CI fails if they’re stale. When I wanted a dealbreaker level, my agent asked James’s agent on the board, the contract changed first, and my Profile change shipped on top of it.

• Every place has a URL

  The open listing, the view and the sort live in the URL, so a listing can be shared, a reload keeps your place and the back button closes the modal. The places on either side load in the background, so the arrows feel instant.

• Maps and voice with no keys

  MapLibre and OpenFreeMap need no account. Next doesn’t bundle MapLibre’s web worker, so a build step copies it in, and a browser without WebGL2 gets a note instead of a broken page. The mic records in the browser, and Whisper turns it into text right there.

• The newest Next broke every page

  Next 16.4 added a file the Cloudflare adapter doesn’t bundle, so every page returned a 500. I pinned 16.3.8 until the adapter catches up.

• One site blocked every crawler

  Apartments.com’s bot protection stopped every approach we tried, so it isn’t a source.

• Search cards leave out the facts

  Zillow’s results have no fees, pets or descriptions, so its crawler reads each listing page once and keeps it for 14 days, and Fred reads a page itself when it needs to know.

• The crawlers ran the wrong search

  We found they’d been running a placeholder search instead of the one in the app. Now Fred starts them on demand, for your own search.

ToolsThe tools Fred can call, in cloud/src/agent.

25

EvalsScripted conversations in evals/fred.ts that check what Fred did, not just what it said.

15

Listing sitesStreetEasy, Zillow, RentHop, Zumper, Compass, Nooklyn and Craigslist, plus big landlords’ own sites.

7

My commitsMy commits to the Fred repo on October 11, 2026, not counting merges.

37

You can try all of it in the live demo, which signs you in as a fresh guest with an example search.

• Read the full listing page on every site, not only Zillow, so fewer facts say Unknown.
• Run the evals in CI and block a deploy when one fails.
• Send alerts by text as well as email.
• Open it to renters beyond the two of us, with real accounts instead of two logins and a demo.