Recalled Rides

Recalled Rides exposes a public, read-only Model Context Protocol server over Streamable HTTP at:

https://recalledrides.com/mcp

It provides four tools backed by the same D1 recall database as the website:

• search_vehicles — find make/model/year combinations with recalls
• list_models — list recalled models for a manufacturer
• get_vehicle_recalls — retrieve plain-English recall details for one vehicle
• get_recall_campaign — retrieve an NHTSA campaign and its affected vehicles

For local development, run npm run dev and configure an MCP client to use http://localhost:8787/mcp. The server is intentionally read-only and does not expose ingestion, enrichment, alert, or other administrative operations.

Cloudflare-native vehicle recall search. Aggregates NHTSA safety recall data, enriches it with LLM plain-English summaries, and serves thousands of SEO-optimized pages from the edge.

Stack: Hono → D1 (SQLite) → Cache API → Workflows (ingestion + enrichment) → Agents SDK (admin orchestration)

# 1. Install dependencies
npm install

# 2. Create your local secrets
cp .dev.vars.example .dev.vars
# Edit .dev.vars and set ADMIN_TOKEN (generate with: openssl rand -hex 16)

# 3. Create D1 database (first time only)
npx wrangler d1 create recall-radar-db
# Copy the database_id into wrangler.jsonc → d1_databases[0].database_id

# 4. Apply database migrations
npm run db:generate
npm run db:migrate:local

# 5. Start dev server
npm run dev

Open http://localhost:8787 and the admin dashboard at http://localhost:8787/admin.

NHTSA vPIC API ──→ IngestionWorkflow ──→ D1 (raw recall data)
                                              │
                                              ▼
                                    EnrichmentWorkflow ──→ Workers AI ──→ D1 (enriched text)
                                              │
D1 ──→ Hono routes ──→ HTML templates ──→ Cache API ──→ HTTP response

# 1. Verify everything compiles
npm run typecheck

# 2. Generate and apply database migrations
npm run db:generate
npm run db:migrate:remote

# 3. Set production secrets
npx wrangler secret put ADMIN_TOKEN

# 4. Deploy
npm run deploy

# Trigger a test ingestion (single make)
curl -X POST https://recalledrides.com/api/admin/ingest \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"mode": "single-make", "targetMake": "TOYOTA", "yearStart": 2023, "yearEnd": 2024}'

# Check status
curl https://recalledrides.com/api/admin/status \
  -H "Authorization: Bearer $ADMIN_TOKEN"

# Verify pages render
curl -s https://recalledrides.com/ | head -20
curl -s https://recalledrides.com/toyota | head -20
curl -s https://recalledrides.com/toyota/camry | head -20
curl -s https://recalledrides.com/toyota/camry/2023 | head -20

# After ingestion completes, trigger enrichment
curl -X POST https://recalledrides.com/api/admin/enrich \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"batchSize": 50}'

# Submit sitemap to Google
curl https://recalledrides.com/sitemap.xml

One weekly trigger (Monday 02:00 UTC): delta ingestion (skips rows checked within last 6 days), which kicks off LLM enrichment as its final step.

• Upsert idempotency: All writes use ON CONFLICT DO UPDATE — safe to re-run
• Raw vs enriched: recalls table keeps government text in *_raw columns; LLM output in *_enriched. Templates fall back to raw when enriched is null.
• Page cache keys: page:<type>:<slug> pattern (stored in the Cache API under /__cache/<key> URLs), TTLs vary by page type
• Admin auth: Authorization: Bearer <ADMIN_TOKEN> on all /api/admin/* routes
• Severity: Auto-classified at ingest from component keywords (CRITICAL → LOW)