| name | car-finder-il |
| description | Find a good second-hand car in Israel (Yad2) and get pinged on your phone only when something genuinely worth it shows up. Reads listings via an Apify actor (no login, no browser, reads-only), scores each car against its OWN peer cohort (price vs comparable listings, hand / יד, owner type, km-per-year, photos), and on a schedule pushes ONLY new matches and real price drops to WhatsApp (wa-ai), Telegram or ntfy. Use whenever the user wants to: search used cars, "find me a car", set buying criteria, watch the market, get alerts smarter than Yad2's, or compare a specific car's price to similar ones. Triggers on: רכב יד שנייה, חיפוש רכב, קניית רכב, יד2 רכב, התראות רכב, "find me a used car", "watch for cars", "is this car a good price". Reads-only — it finds and ranks, it does not post, bump, or buy. |
Car Finder IL (חיפוש רכב יד שנייה)
Turns the noise of Yad2 into a short, ranked shortlist and a phone ping. It does
the one thing Yad2's own alerts don't: it only tells you about a car when the car
is actually good — priced under its peers, low hand, private owner, sane km.
What this CAN and CANNOT do
CAN (public, reads-only): read live Yad2 vehicle listings — price, make,
model, year, km, hand (יד), owner type, photo count, link — through a maintained
Apify actor that absorbs Yad2's anti-bot. Score them against their own cohort.
Remember what it already showed you. Push new hits to your phone on a schedule.
CANNOT: publish, bump (הקפצה), edit, message sellers, or buy. There is no
posting API on any Israeli car board; this skill stays firmly on the read side
on purpose. It's a finder, not a bot, so it can't get your account flagged.
Baseline it has to beat: Yad2's saved-search alerts are free but fire on
every crude filter match. The whole point here is the scoring — pinging you
only when a listing beats the median of comparable cars, not on everything.
No מחירון API — so we benchmark honestly
לוי יצחק has no public API, so the skill doesn't fake one. Instead each car is
scored against its own cohort — same make+model, year within ±1, pulled in
the same fetch. "12% under comparable listings" is self-contained and true.
If you have a personal target price per model, put it in your criteria and the
hard filters enforce it.
The engine
All logic is stdlib-only Python under scripts/ — no install. Run with bash;
add --json to search for machine-readable output.
python3 scripts/carfinder.py search examples/criteria.example.json --source mock --lang he
python3 scripts/carfinder.py search my_criteria.json --lang he
python3 scripts/carfinder.py run my_criteria.json
python3 scripts/carfinder.py init-db
| File | Does |
|---|
scripts/carfinder.py | CLI, EN/HE rendering, orchestration |
scripts/source.py | fetch listings (Apify adapter + mock), normalise |
scripts/score.py | hard filters + cohort-relative weighted score |
scripts/store.py | Supabase dedupe (PostgREST); local-file fallback offline |
scripts/notify.py | push to wa-ai / Telegram / ntfy / stdout |
Criteria file
A small JSON the user edits — see examples/criteria.example.json. Hard filters
(price_max, year_min, km_max, hand_max, owner_type, make) gate what's
even considered; weights tune the score; min_score sets the ping threshold;
label and lang shape the message. Keep one file per search; the id field
is the dedupe key.
How to drive this as the assistant
- If the user describes what they want ("family SUV, private owner, under 90k"),
write them a criteria JSON — don't make them learn the schema.
- Run
search --source mock first to show the shape, then search on real data
once their Apify keys are in. Always surface the reasons, not just a score.
- Only suggest the cron + Supabase + phone-push (the
run path) once they like
the shortlist. It's optional; search alone is fully useful.
- Stay honest: this finds and ranks. For posting/selling, that's a different,
browser-driven problem — don't imply this can buy or message anyone.
