An agent that runs a rental search end to end: sources listings, filters and ranks them, calls the listing contact with a Dial AI voice agent, extracts cited facts with OpenAI (or Anthropic), and collects viewing offers for the renter to confirm.
Phase 1 vertical slice of the build plan.
The current integration includes the household Next.js app from app-ui-form-and-login and the Python agent worker from ui-refresh. See agent integration and deployment for the shared API, required settings, live search evidence, and voice support.
sources (citation-backed web search · public property pages · craigslist · Gmail alerts)
└─> collect: normalize, pull phone from text, dedupe, filter, score
└─> cross-search: address/contact corroboration, conflicting terms, comparable rents
└─> call (Dial POST /api/v1/calls, poll GET /api/v1/calls/{id})
└─> extract (OpenAI/Anthropic: facts + exact transcript evidence)
└─> exact listing verification → confirmed tours or unbooked offers
python3 -m venv .venv && .venv/bin/pip install -r requirements-dev.txt
cp .env.example .env # defaults: dry-run calls + mock listings
.venv/bin/uvicorn househunt.web.app:app --env-file .env --reload
.venv/bin/python -m pytest -qOpen http://localhost:8000, start a search, and watch it move through sourcing → calling → done.
| Setting | Values | Notes |
|---|---|---|
CALL_MODE |
dry_run · approve · auto |
approve = a "Call" button per listing; auto = top MAX_AUTO_CALLS dialed automatically |
MOCK_SOURCES |
1 · 0 |
Mock listings have fake 555 numbers, so real dialing is disabled while this is 1 |
dial auth login you@example.com # email OTP, then phone verification
dial number list # your from-numberPut the API key from ~/.local/share/dial/auth.json in DIAL_API_KEY and the number in
DIAL_FROM_NUMBER. Free Dial accounts cap calls at 5 minutes and 2 concurrent calls; the app
respects both.
render.yaml is a blueprint: Render → New → Blueprint → this repo. It builds the
Dockerfile, mounts a 1 GB disk at /data for SQLite, and generates APP_PASSWORD
(HTTP basic auth, any username). Set DIAL_API_KEY, DIAL_FROM_NUMBER, NOTIFY_NUMBER,
and OPENAI_API_KEY (or ANTHROPIC_API_KEY) in the dashboard. Any Docker host works the same way.
The current web app is Python/FastAPI on the team's ui-refresh branch. Its existing Call button now runs the native Python caller; no separate Node service is required. caller/ remains the standalone TypeScript shared-contract harness from the earlier brief. lib/types.ts is unchanged.
For live calls set CALL_MODE=approve, MOCK_SOURCES=0, DIAL_API_KEY, DIAL_FROM_NUMBER, and OPENAI_API_KEY. Use APP_PASSWORD when exposing the app. Keep .env and data/ private and out of Git. Run a single web worker with a persistent SQLite disk. The configured calling timezone defaults to America/Los_Angeles.
The default call cap is 180 seconds. Calls are serialized, fictional 555 numbers are rejected, and each contact defaults to one attempt per day across searches (MAX_CALLS_PER_NUMBER_PER_DAY). Duplicate clicks reuse the saved attempt rather than creating a new call. Resume saved call polls or re-extracts the existing call; ambiguous placements use the same persisted Dial idempotency key. Interrupted work enters attention after a process restart and needs an explicit resume.
Live extraction uses OpenAI when OPENAI_API_KEY exists, otherwise Anthropic. It never falls back to keyword guesses. The raw response and transcript are saved before extraction; provider errors are shown without response bodies or credentials. Opt-outs persist even if extraction fails. Dry runs are explicitly labeled synthetic and never invoke either model provider.
calls.outcome includes checks (real/matches/fit/tour with {callId,t,quote} evidence), transcript_turns, offered_windows, verification_status, listing_verified, confirmed_tour, booking_warning, and mode, while preserving existing summary fields. The green listing check requires the contact to confirm the exact address/unit, authority, availability, advertised rent and bed/bath counts. Financial eligibility remains separate. The agent can book a concrete tour within freshly checked household calendar availability after both sides explicitly confirm it; missing calendars leave it collecting offers. Only confirmed bookings create tours, viewing-tracker entries and calendar events. Calendar retries are idempotent and never redial. See the integration contract.
Dial's dashboard provides recordings through expiring signed links; the documented public API does not expose a recording-download endpoint. For demo replay, download the call's WAV from Dial → Voice Calls and import it with:
.venv/bin/python -m househunt.outreach.recording CALL_ID /path/to/download.wavCALL_ID is the local SQLite call ID. Imported audio is stored beside the database with owner-only permissions and served by the app's authenticated /calls/:id/recording route, including range requests for seeking. Timestamp links open the matching transcript line and seek the player. The URL stays null until a recording is imported; the app never exposes Dial credentials or signed storage links. Real call data remains local; the committed fixture in mocks/ is synthetic.
| Source | Status |
|---|---|
| Craigslist | Live: no-JS search page + each posting. Phone numbers only when written in the description |
| Web discovery | OpenAI web search citations, followed by public-page reads and exact-quote fact validation. One level of observed inventory links, up to two discovery rounds |
| Zillow / other portals | Public pages can be discovered by web search; blocked pages are reported, not bypassed. Gmail Zillow alerts and manual imports remain available |
| Facebook Marketplace | No logged-in automated browsing. Manual import supported; mock fixture available in mock mode |
househunt/
models.py Listing, SearchCriteria, statuses
sources/ one module per site, each with fetch() + mock_results()
collect.py normalize / dedupe / filter / score
outreach/
script.py exact-listing verification + authorized tour-booking prompt
booking.py fresh availability, confirmed tours and calendar retries
dial.py Dial REST client + FakeDialClient for dry runs
extract.py transcript -> cited CallSummary via OpenAI/Anthropic
verification.py exact-quote, timestamp, requirement and tour-window checks
research/ bounded web discovery, page evidence and rental screening
platform.py SearchBrief polling and AgentRun publishing to Next.js
pipeline.py orchestration (parallel research, serialized calls)
tours.py .ics invites
db.py SQLite
web/ FastAPI app + Jinja templates
Autonomous outreach now uses a durable office-hours queue, evidence-backed voicemail handling, and inbound callback capture. Enable the worker flags and connect both app/worker URLs as described in the outreach deployment contract. Scheduled times and their hours sources appear in both dashboards. Live calls retain the exact search/listing context and only verify facts supported by contact quotes.