ai apartment hunting
Apartment hunting is a reading problem wearing a browsing costume. The work is not clicking through listings; it is holding forty of them in your head against a dozen criteria - budget, commute, floor, pets, lease length, which "cozy" means small and which "needs love" means broken - and doing it again tomorrow when thirty new ones appear. That is judgment applied to prose at volume, which is precisely the shape of task an AI agent with a real browser is good at, and precisely the shape a scripted scraper is bad at, because the facts live in free text and photos, not in stable fields.
This page covers what the agent genuinely does for a search like this, a
concrete workflow with the aihawk CLI, and the boundaries stated up
front rather than in the fine print - because listing portals have terms
too, and an agent does not exempt you from them.
Listing sites have terms of use, and many restrict automation. The big portals - Zillow, Apartments.com, Idealista, Rightmove, ImmobilienScout24, whichever ones serve your market - publish terms, and prohibitions on scraping and automated access are common; Rightmove's site states flatly that it "prohibits the scraping of its content." Read the terms of the portal you actually use before pointing an agent at it. An agent reads a page the way a person does, one listing at a time under your instruction - but the terms are the site's call, not this wiki's.
Volume is still volume. A person checking twenty listings a day is a reader; a scheduled sweep pulling five hundred listings an hour is a scraper, whatever software does the pulling. The signals that get agents blocked anywhere - rate, rhythm, fingerprint - apply on listing sites too, and the map for them is why does my AI agent get blocked?. Keep the ask human-sized and most of that page never becomes relevant to you.
The agent reads; you contact. Inquiries to landlords and agents are messages sent in your name about a place you might live. The agent can draft one from a listing's details; a human sends it. Auto-sending inquiries at volume is the rental-market version of the spray pattern, and it burns the same thing every volume play burns: your credibility with the people you need to say yes.
Reads listings against your criteria, not just filters. Portal filters stop at structured fields. Your real criteria do not: "ground floor only if there is a garden," "top floor only with an elevator," "no north-facing," "landlord-managed preferred." An agent holds the full criteria list and reads each listing's prose and captions against it, returning fits, near-misses with the reason, and the disqualified with the disqualifier named - so you can audit its judgment instead of trusting it.
Compares across sites that do not compare themselves. Markets are fragmented across portals with different layouts and vocabularies. The agent does not care; a listing page is a listing page. Ask it to check the same neighborhood on two portals and reconcile duplicates - the same flat posted twice at different rents is a genuinely useful find.
Extracts to something you can track. A search generates state: what you saw, where, at what price, its status. Have the agent emit each session's findings as CSV rows and you have a running spreadsheet of the market - the mechanics and the honest failure modes of that pattern are covered in extracting data to a CSV with an AI agent.
Watches for new listings - with judgment, on a sane schedule. The recurring question in a hot market is "anything new today that fits?" That is a monitoring task where the condition needs reading, which is exactly when an agent earns its per-run cost over a plain diff tool - the boundary the monitoring page draws in detail, including when the plain tool wins. Daily is a search; every five minutes is a bill and a signature.
Reading listings against criteria is judgment work, so it runs where the
model is: uvx aihawk ui, or your assistant with AIHawk's browser attached.
A worked instruction to paste, with the portal URL being whatever
search-results page you have already set up by hand:
Go to
<your saved search URL>. Read the first 15 listings. My criteria: max 1600/month, 2 rooms, pets allowed, available within 60 days, no ground floor without outdoor space. For each listing output one line: title, price, rooms, floor, verdict FIT/NEAR/NO with the deciding reason. Read the listing text, do not trust the summary card. If a fact is not stated, write UNKNOWN, do not guess.
The prompt carries the craft, and three parts of it matter more than the rest. The criteria are explicit and closed - the agent judges against your list, not its taste. The output is one line per listing with the reason attached - auditable at a glance. And "UNKNOWN, do not guess" is load-bearing: listings omit exactly the facts that matter, and an agent that fills gaps optimistically is worse than none, because a wrong "pets allowed" costs you a viewing trip.
For the recurring half - "anything new today?" - split the work the way
the monitoring page does: a
scheduled script on the same engine captures the listing count or the newest
title each morning, and when that signal moves you bring the judgment prompt
above to the agent. Once a day, or twice in a genuinely fast market. For
interactive sessions - "open the third FIT and tell me what
the photos show about the kitchen" - uvx aihawk ui gives you the same
agent beside a live browser view, and if you already use Claude Code or
Claude Desktop, the same browser attaches to your assistant instead
(the setup page has the one-liner).
Two practical notes from the field. First, listing portals are heavy, banner-laden pages; a stronger model earns its cost here more than on simple pages, and which model to use with AIHawk covers that trade. Second, keep the geography honest: if you run through a proxy for other work, a search "from" the wrong country gets you the wrong inventory and prices - plain home connection is the right default for a local search.
The agent compresses the reading. It does not - and should not - replace the parts where the stakes live: viewing the place, judging the neighborhood at 10pm rather than in the photos, reading the lease, smelling the damp the wide-angle lens cropped out. Treat its output as a briefing, verify anything that costs money to believe (a "FIT" is a claim, not a fact - the listing may have lied, or the agent misread), and never send money or documents based on a listing nobody has visited. Rental fraud predates AI and an agent does not detect it for you; a deal that reads too good in the agent's summary reads that way because it is.
Can an AI agent find me an apartment? It can read the market for you: screen listings against real criteria, compare across portals, track what is new, and draft inquiries. Choosing, viewing, and signing stay human - as does hitting send on the first message.
Is it allowed to use an agent on listing sites? The portal's terms decide, and several restrict automated access - Rightmove states it prohibits scraping outright. Read the terms of the site you use; a human-paced, human-supervised session is a different thing from a scraping operation, but the line is drawn by the site, not by your tooling.
How is this different from setting up portal alerts? Alerts fire on structured filters; the agent judges prose criteria the filters cannot express, reconciles across portals, and explains each verdict. Use both: alerts for speed, the agent for judgment.
What does it cost per run? A browser session plus a handful of model calls - cents per check on the default model, roughly a minute of wall clock. Fine daily; wasteful as a rapid poller, and a plain diff monitor is the better rapid poller anyway.
Will the agent get blocked by listing sites? At human pace with the project's browser, usually not - and if it does, work the layers in order on the blocking page before blaming any one of them. A sweep at scraper volume deserves the block, and this page is not the recipe for one.
Can it fill out rental application forms too? Yes, with the same discipline as any form: draft from your real information, stop at anything binding, human reviews and submits. See getting an AI agent to fill out forms.
All retrieved 2026-09-03.
- The AIHawk README, for the
uvx aihawk uicommand, the MCP path for assistants, and the profile and proxy behavior (updated for aihawk 0.3.0, which removed thedosubcommand). - Rightmove's terms-of-use page, for its stated prohibition on scraping its content, cited as the concrete example that listing portals restrict automated access.
- Portal names (Zillow, Apartments.com, Idealista, ImmobilienScout24) are used as examples of the category only; no claim is made here about any individual site's terms beyond the one quoted above.
See also: monitoring a page for changes with an AI agent for the recurring-check mechanics this page leans on, extracting data to a CSV with an AI agent for turning sessions into a tracked spreadsheet, which model to use with AIHawk for the model trade-off on heavy pages, and why does my AI agent get blocked? for when a portal pushes back.
From the AIHawk wiki. The agent reads the listings; you still take the viewing - and the flat with the suspiciously wide-angle photos is still small.