Inspiration

Moving out means a pile of small chores, and every missed one comes out of your security deposit:

$$D_{\text{back}} = D - \sum_i c_i \cdot \mathbb{1}[\text{task } i \text{ missed}]$$

Here $c_i$ is the deduction for each missed task. Vacate’s agents handle those tasks so that sum goes to zero.

What it does 📞 Calls your landlord: an AI voice agent asks for the top 3 move-out items and saves them as your checklist. 🤖 Runs from Claude via MCP: 30+ tools let you manage your whole move from a chat. 📘 Posts to Facebook through Muse: Claude prepares the listing, and Muse posts it from your account. 💬 Sells on Telegram: a negotiation agent haggles with buyers and books pickups. It never goes below your minimum price. 📄 Reads your lease and compares cleaners and movers, with page references and 3 options each. ✅ You approve every call, post and deal. How we built it

Next.js 16 on Vercel; Supabase (Auth, Postgres with row-level security, Storage); Claude via OpenRouter; Bland AI for calls; an MCP server with OAuth; a Muse bridge; and the Telegram Bot API.

Challenges we ran into Making an AI negotiator safe: the model writes the replies, and code enforces the price floor, address privacy and pickup deadline. Posting to Facebook with no public API, solved through Muse. Keeping one user identity across the website, Claude and Muse. Accomplishments that we’re proud of

A real landlord call that becomes a checklist, and one set of tools serving the website, Claude and Muse. All live, built in one day by four people.

What we learned

Let the model write; let code decide. A clear approval step is what makes people trust an agent.

What’s next for Vacate

Real cleaner and mover booking, more marketplaces, and move-in condition photos for disputing deposit deductions.