carsearch — Canadian used-car search, monitoring & market-analysis platform

carsearch turns Canadian marketplace listings into a persistent, searchable, historical database and exposes it to Claude Code (or any agent) over MCP, plus a full CLI. It collects real listings from AutoTrader.ca (more providers pluggable), keeps every price/status change as history, deduplicates physical vehicles across listings, geocodes properly (real distances), extracts risk signals from descriptions, computes comparable-price statistics, scores project cars under configurable profiles, detects deals, reruns saved searches on a schedule and raises alerts.

Ask Claude things like:

• Find me every manual BMW E90 328i under $8,000 within 800 km of Toronto, prefer private sellers, reject rust buckets, rank by project-car value.
• What interesting project cars under $7k appeared in Ontario in the last 48 hours?
• Is this $6,200 Audi TT actually cheap relative to similar listings we've observed?
• Show me every price drop > 10% this week. / Cars sitting 30+ days where the seller already cut the price twice.

Status per phase, tested commands and known issues: PROJECT_STATUS.md.

---

Architecture (short)

providers/ (autotrader ✓, cargurus/kijiji/facebook interface+probe)  →  VehicleListing (canonical)
   → collectors/ (checkpointed ProviderRun, ingest w/ snapshots+events, raw payloads)
   → dedupe/ (VIN → weighted soft signals → vehicles + candidates)  → geo/ (geocode, haversine)
   → SQLite/PostgreSQL (SQLAlchemy)  → search/  analysis/ (description, comps, deals, market)
   → scoring/ (profiles + model knowledge JSON)  → monitoring/ (saved searches, scheduler) → notifications/
   → cli/ (carsearch …)  and  mcp/ (23 tools)

Details: docs/architecture.md, docs/providers.md, docs/schema.md.

Installation

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"            # add ".[postgres]" for PostgreSQL
cp .env.example .env               # optional; SQLite in ./data works out of the box
carsearch db-init                  # creates/upgrades data/carsearch.db

Python ≥ 3.11 (developed on 3.14). No browser, no API keys required for AutoTrader.

Database setup

• Default: sqlite:///./data/carsearch.db (WAL). Nothing to do.
• PostgreSQL: CARSEARCH_DATABASE_URL=postgresql+psycopg://user:pass@host/carsearch in .env, pip install -e ".[postgres]", carsearch db-init. Schema/migrations in src/carsearch/database/ (versioned, recorded in schema_migrations).

Crawling (collecting real data)

# targeted, with detail pages for new listings (VIN, exact coordinates, drivetrain, created date)
carsearch crawl autotrader --make BMW --model "3 Series" --transmission manual --max-price 8000 --min-year 2006 --max-year 2013 --details new
# regional bulk (20 listings/page, ~1.5 s/page, no detail pages)
carsearch crawl autotrader --max-price 7000 --location Toronto --radius-km 500 --max-pages 200 --details none
# province scope (centroid + covering radius server-side, exact province locally)
carsearch crawl autotrader --province ON --transmission manual --max-price 6000 --seller-type private
# incremental refresh: stop after 3 consecutive pages with nothing new/changed
carsearch crawl autotrader --transmission manual --max-price 8000 --stop-when-seen 3
# resume an interrupted run
carsearch crawl autotrader --resume 12
# re-check listings not seen for 3 days (marks removed on 404); fetch missing detail pages; geocode backfill
carsearch refresh autotrader --older-than-days 3 --limit 200
carsearch details autotrader --limit 200
carsearch geocode --limit 60
scripts/bootstrap_crawl.sh   # example multi-band bootstrap used to seed the DB

AutoTrader caps a query at 200 pages (4000 listings) → split by price band/region for full coverage (truncated=True in the run summary tells you when).

Searching

carsearch search --make BMW --model 328i --max-price 7000 --transmission manual --location Toronto --radius-km 800 --seller-type private
carsearch search --generation E90 --keywords '"one owner" -salvage' --province ON,QC --sort price_asc
carsearch new --since 48h --province ON --max-price 7000
carsearch price-drops --min 10 --since 7d
carsearch deals --province ON --max-price 7000 --profile project_car_enthusiast
carsearch deals --location Toronto --radius-km 500 --transmission manual --max-price 7000 --since 7d \
    --profile project_car_enthusiast --sort-by project --exclude-flags salvage_title,frame_rust,doesnt_run   # "best project cars" ranking
carsearch listing 42            # full record + price history + description flags
carsearch history 42            # snapshots + change events
carsearch comps 42              # comparable-price analysis with methodology
carsearch score 42 --profile project_car_enthusiast
carsearch analyze 42            # everything above + deal signals in one JSON
carsearch market --metric manual_premium --make Subaru --model WRX
carsearch market --metric price_by_mileage --generation E90 --province ON
carsearch market --metric cheapening --days 90
carsearch status [--check-providers]  ·  carsearch runs  ·  carsearch alerts

Every command accepts --json for machine-readable output.

Monitoring / saved searches / alerts

carsearch saved seed                       # the four example searches from the brief
carsearch saved create "manual E90 328i under $8k" '{"make":"BMW","model":"3 Series","min_year":2006,"max_year":2011,"max_price":8000,"transmission":"manual"}' \
    --interval-minutes 240 --alert-rules '{"price_drop_pct":10,"target_price":6500,"min_deal_score":65}' --profile project_car_enthusiast
carsearch saved run "manual E90 328i under $8k"   # incremental crawl → new listings + alerts
carsearch schedule                          # long-running loop: runs due searches, maintenance
carsearch schedule --once                   # one pass (cron-friendly)

Alerts go to console, data/alerts.jsonl and an optional webhook (CARSEARCH_WEBHOOK_URL); new sinks = subclass notifications.alerts.AlertSink. Every alert carries reasons[] (and risks).

MCP server

carsearch serve-mcp                # stdio (default)  |  --transport streamable-http --port 8765

Tools: search_cars, search_live_marketplace, get_listing, get_vehicle, get_new_listings, get_price_history, get_listing_changes, get_comparables, compare_cars, find_deals, find_price_drops, find_long_sitting_listings, analyze_listing, score_listing, analyze_market, list_saved_searches, create_saved_search, run_saved_search, delete_saved_search, get_alerts, get_status, get_generation_codes, analyze_listing_images. All return structured JSON with pagination (limit/offset/next_offset); heavy fields (description, photos, raw) are opt-in.

Connect Claude Code

claude mcp add carsearch -- /ABSOLUTE/PATH/autotrader/.venv/bin/carsearch serve-mcp
# or in .mcp.json:
{ "mcpServers": { "carsearch": { "command": "/ABSOLUTE/PATH/autotrader/.venv/bin/carsearch", "args": ["serve-mcp"] } } }

Then: "Find me the best manual project cars under $7,000 CAD within 500 km of Toronto that appeared in the last week. Compare them against historical prices, flag major mechanical/rust risks from the listings, and rank the top 10." — Claude will typically call search_cars/get_new_listings (or search_live_marketplace for fresh data), then analyze_listing/compare_cars.

How checkpoints work

Each crawl is a provider_runs row. After every page the listings are committed and last_checkpoint.next_page is advanced; per-listing failures are logged in errors and skipped. Detail enrichment commits per listing (detail_fetched_at). A run that dies is failed/interrupted and carsearch crawl --resume <id> continues from next_page (already fetched pages come from the local http_cache). Saved-search state lives in the DB. See docs/architecture.md.

How to add a provider

Implement BaseProvider in src/carsearch/providers/<name>/, map to VehicleListing in a parser.py (keep raw), register it, add fixtures + tests, document it in docs/providers.md. Nothing else changes.

Tests

pytest -q          # 42 tests: normalization, parser fixtures, ingest/price history, dedupe,
                   # geo/search, analysis (flags/comps/scoring/deals/market), crawler checkpoints, MCP tools

No test touches the network (real payload fixtures + a fake provider).

Known limitations

• Providers: only AutoTrader.ca collects. CarGurus (DataDome), Kijiji (edge 403 / kijijiautos.ca gone) and Facebook (login) are interfaces with live health probes; no anti-bot or login circumvention is attempted (see docs/providers.md).
• Asking prices, not sale prices. Disappearance ≠ sold; we record not_observed/removed.
• Coverage depends on what you crawl (200-page cap per query; split by bands). "Appeared in the last N days" uses the marketplace creation date when known (detail page) else our first observation.
• Coordinates: exact for listings with detail pages; otherwise city-level (seed table / learned centroids / marketplace resolver). Distance filters skip listings without coordinates.
• Scoring, flags and deal scores are heuristics with stated evidence and configurable rules (src/carsearch/config/*.json); they are inputs to reasoning, not verdicts. Regex flags can misfire on unusual phrasing (evidence is always returned).
• Comparables use progressive relaxation; small samples are flagged (warnings, confidence).
• Image analysis ships hashing/caching and a pluggable vision hook (Anthropic backend included, off by default); no vision model runs unless configured.
• SQLite is fine for hundreds of thousands of listings; switch to PostgreSQL for concurrent writers.