🎯 Sublet Sniper

Your rental agent for San Francisco. Say what you want; it searches, blocks scams, emails and phones the landlord, negotiates, and books the viewing. You just show up.

Live: https://sublet-sniper.vercel.app · App: https://sublet-sniper.vercel.app/app (Continue as guest) · MCP: https://sublet-sniper.vercel.app/api/mcp

Who it's for: anyone moving to SF for 1 to 6 months (interns, new hires, founders in a batch, visiting researchers) who has to rent sight unseen. That gap lives on Craigslist sublets, and it's exactly who scammers target: "I'm abroad, Zelle me the deposit." In August 2026 a renter wired $18,600 for a Duboce Triangle 2BR copied from a real Zillow listing (SF Standard, KQED, ABC7).

What's different: listing sites show you listings. Sublet Sniper acts for you:

• Scam X-ray on every listing: deterministic, explainable score (price vs SF median, payment red flags, duplicate text, reused photos via pHash, upfront > 2x rent per CA AB 12).
• Calls the landlord with an ElevenLabs voice agent (Twilio number): refuses any payment before a viewing, negotiates on a fixed ladder inside your budget, books the viewing mid-call. Live captions stream on screen while the call happens.
• Emails the landlord from its own AgentMail inbox and keeps negotiating by email.
• One human step: pick the listing. Then add the viewing to your calendar.
• Agent-native: the same powers over MCP (search_sublets, check_listing_scam, contact_landlord, get_deals, ask_agent), plus llms.txt.
• Multi-tenant: Supabase auth (magic link or guest) with row-level security.

Built for the Puentes hackathon (SF, 2026-10-24). Demo script: docs/DEMO.md.

What it does

Tool	What it does
search_listings	Live Craigslist SF (sub, apa, roo), polite (≥3 s between requests, browser UA, 10 min cache), falls back per category to a 765-listing real snapshot; plus X posts snapshot and labeled demo fixtures
normalize	Heuristics (price period, totals vs nights, type, furnished, dates, in-SF) + optional LLM pass validated with zod
rank	Budget (hard cap 110%), type, furnished, date overlap, haversine distance to named places; scam light penalizes
score_scam	Green / yellow / red with reasons: wire/Zelle/gift card/out-of-country keywords, price < 60% of the neighborhood median (Apartment List / Zumper / Craigslist medians), duplicate text (shingle Jaccard), duplicate photos (pHash Hamming ≤ 6), upfront > 2× rent (CA AB 12), new X accounts, empty bodies
draft_contact	Drafts the first message into the approvals queue. Craigslist: approving copies the text and opens the listing, the human clicks Reply (captcha). Email only to allowlisted demo addresses
landlord loop	AgentMail inbox: answers from the renter profile, counters inside budget, asks for a video walkthrough / leaseholder proof, proposes slots; refuses and flags deposit-before-viewing, Zelle, wire, crypto, "out of the country", keys by mail
propose_slots / book_viewing	Google Calendar freeBusy + events.insert (booking needs your approval) or mock calendar
ask_owner_decision	Goes to the approvals queue; the only time you get pinged

Dashboard: chat + voice (🎤 Web Speech API in/out, optional ElevenLabs widget), ranked cards with traffic lights, flagged tab, a thread per landlord, the agent's action timeline, and the approvals queue.

Architecture

flowchart LR
  U[You: voice / text] -->|Web Speech API or ElevenLabs| UI[Next.js dashboard]
  UI -->|/api/agent| AG[Agent loop<br/>Claude → xAI → OpenAI → offline mock]
  AG --> T1[search_listings]
  T1 --> CL[(Craigslist live<br/>polite fetch)]
  T1 --> SNAP[(Snapshot: 765 CL listings,<br/>X posts, pHashes, baselines)]
  T1 --> DEMO[(DEMO fixtures:<br/>accomplice + $18.6k Duboce scam)]
  T1 --> N[normalize] --> S[score_scam] --> R[rank]
  AG --> DC[draft_contact] --> AQ[[Approvals queue]]
  AQ -->|Craigslist| HUMAN[Human clicks Reply]
  AQ -->|demo email allowlist| AM[AgentMail inbox]
  AM -->|webhook / poll| LL[Landlord loop<br/>safety rules first, LLM wording]
  LL --> AM
  LL -->|red flags| AQ
  LL --> CAL[Google Calendar<br/>freeBusy + insert]
  AG --> ST[(In-memory store:<br/>timeline, threads, approvals)] --> UI

Setup

git clone https://github.com/ariasnico/sublet-sniper && cd sublet-sniper
npm install
cp .env.example .env.local   # optional: everything works with zero keys
npm run dev                  # http://localhost:3000
npm test                     # vitest on real snapshot data
npm run build

Modes (shown as pills in the header): with no keys you get the deterministic offline agent (Spanish/English), live Craigslist (or CRAIGSLIST_LIVE=0 for snapshot), mock mail with "Simulate DEMO landlord" buttons, mock calendar, and browser voice. Add keys to go live, see .env.example for where to get each one.

Deploy to Vercel: import the repo, add env vars, done. Without Supabase env vars the app is single-tenant (one in-memory session per server instance).

Accounts (multi-tenant, Supabase): with NEXT_PUBLIC_SUPABASE_URL + NEXT_PUBLIC_SUPABASE_PUBLISHABLE_KEY (or _ANON_KEY) + SUPABASE_SERVICE_ROLE_KEY set, /app requires sign-in (/login: email magic link, or Continue as guest = Supabase anonymous sign-in, falling back to a throwaway confirmed user if anonymous sign-ins are off). Each user's whole session lives in public.sessions (one jsonb row, RLS: own row only); API routes run inside an AsyncLocalStorage tenant context (src/lib/tenant.ts, src/lib/session.ts), so state() is per user. The shared AgentMail inbox is split by a #<8-hex user id> tag on subjects plus public.thread_index; calls are routed by public.call_index and the user_id dynamic variable (re-run scripts/elevenlabs-setup.mjs to add it to the agent's tools). /api/mcp stays a shared, no-login demo tenant (the global session). AUTH_MODE=off forces single-tenant.

• Schema: supabase/migrations/*.sql, applied with node --env-file=.env.local scripts/migrate.mjs. RLS check: node --env-file=.env.local scripts/rls-smoke.mjs.
• Supabase dashboard (Authentication): Site URL https://sublet-sniper.vercel.app; redirect URLs https://sublet-sniper.vercel.app/**, http://localhost:3005/**; enable Anonymous sign-ins for the guest button. The built-in email sender is rate-limited (a few emails per hour, project-wide), so use Continue as guest for demos or configure custom SMTP.

AgentMail live: create an inbox, set AGENTMAIL_API_KEY, AGENTMAIL_INBOX_ID, DEMO_LANDLORD_EMAIL (your teammate). On Vercel add a webhook to https://<host>/api/mail/webhook?token=<WEBHOOK_TOKEN> for message.received; locally press Check mail.

Use it from your agent

Sublet Sniper is agent-native: the web app and a remote MCP server (Streamable HTTP, stateless) share the same code and safety rules. Endpoint: https://sublet-sniper.vercel.app/api/mcp. Agent-readable overview: /llms.txt.

MCP tool	What it does
search_sublets	{ query?, maxBudget?, type?, furnished?, from?, to?, near? } → top safe matches (id, title, $/mo, neighborhood, scam score + light, url, photo) and the scams it caught
check_listing_scam	{ url } (Craigslist only) → score 0-100, light, structured signals, one-line verdict
contact_landlord	{ listingId } → emails only the allowlisted demo landlord; for real Craigslist listings returns the draft + URL, a human clicks Reply
get_deals	threads with status, negotiation (listed / asked / offered), safety checks, booked viewings with a Google Calendar link
ask_agent	{ message } → the same chat agent as the web UI

Claude Code

claude mcp add --transport http sublet-sniper https://sublet-sniper.vercel.app/api/mcp
# if the deployment sets MCP_TOKEN:
claude mcp add --transport http sublet-sniper https://sublet-sniper.vercel.app/api/mcp --header "Authorization: Bearer $MCP_TOKEN"

Claude Desktop / claude.ai: Settings → Connectors → Add custom connector → name Sublet Sniper, URL https://sublet-sniper.vercel.app/api/mcp.

ChatGPT: Settings → Apps & Connectors → Advanced → enable Developer mode, then Create a connector with URL https://sublet-sniper.vercel.app/api/mcp and authentication No authentication.

Custom connectors in Claude and ChatGPT don't send static bearer headers, so leave MCP_TOKEN unset for them (demo mode). Set MCP_TOKEN to lock the endpoint to clients that can send Authorization: Bearer <token> (Claude Code, MCP Inspector, mcp-remote --header).

Try it locally: npx @modelcontextprotocol/inspector --cli http://localhost:3000/api/mcp --transport http --method tools/list.

2-minute demo

See docs/DEMO.md. In short: ask for a furnished 1BR under $4,000 → ~770 live listings scanned, scams blocked → open a red listing's Scam X-ray → Handle any listing → the agent emails, then phones the landlord (a teammate) → it refuses the Zelle deposit, counters $3,400, agrees $3,600 and books Monday 6 PM, with live captions → booked ticket with your next steps.

Honest limitations

• Craigslist contact needs a human. Replies go through Craigslist's relay behind an hCaptcha, and its ToS forbids robots and unsolicited messages. We never auto-contact Craigslist and never solve captchas; the human clicks Reply. Fetching is low volume (≥3 s between requests, browser UA, cache) and falls back to a snapshot, but scraping is still against Craigslist's ToS. For a real product you'd want a licensed feed or user-initiated browsing.
• Only the allowlisted demo landlord is emailed or called. Real landlords are never contacted by the agent (AI voice calls need consent under TCPA / FCC-24-17). In demo mode, whichever listing you pick is played by the demo landlord.
• X data is a snapshot (live X DMs need a paid dev app and OAuth user token); most "sublet" posts on X are noise.
• Scam scoring is heuristic and explainable, not a guarantee. Duplicate-photo detection only covers images we've hashed.
• Session state is one JSON row per user in Supabase (RLS); live captions use Twilio real-time transcription. Google Calendar uses a refresh token from your own OAuth client (Testing mode).
• Demo data is clearly labeled [DEMO DATA]: the Mission Bay accomplice listing (a teammate plays the landlord) and the Duboce scam fixture modeled on the real case.

Engineering notes

Applied poteto skills: testing (fixture tests on the real snapshot, a full flow test from search to booking to scam refusal), ts-best-practices (discriminated unions with never exhaustiveness, zod at every boundary, no any), react-best-practices (no derived-state effects, single polling effect, parallel tool execution with Promise.all).