How to Book NYC’s Hardest Tables With an AI Assistant (MCP)

“Claude, get me a table for two at Carbone next Friday.” That sentence is closer to working than most people think — but not the way most people think. AI assistants can’t book hard restaurant tables by themselves, and understanding exactly where they fail explains what the working setup looks like. Current as of July 2026.

Why your assistant can’t just book it

Two reasons. The mundane one: the reservation platforms don’t offer official agent connectors. Resy, SevenRooms, and OpenTable are built for humans in apps and browsers, and none of them ships a sanctioned interface an AI assistant can call to make a booking. An agent trying to browse its way through a booking flow is slow, brittle, and often blocked outright.

The structural one is worse: hard tables aren’t bookable at ask-time. The whole problem with a Carbone or a 4 Charles is that when you want it, the book shows nothing — the table appears later, at a timed release or as a random-hour cancellation, and is gone in seconds. Even a perfect agent acting at the moment you asked would find nothing to book. What you actually need isn’t an agent that books now — it’s an agent that can set up a watch that books later, the instant the table exists.

The shape that works: agent → MCP → watcher

The Model Context Protocol (MCP) is the open standard that lets AI assistants call external tools. A reservation MCP server exposes the operations an assistant needs — “watch this restaurant, these dates, party of two, budget X” — and the assistant invokes them mid-conversation. The division of labor is clean: the agent handles intent (“somewhere downtown, Italian, Friday-ish”), the watcher handles the part agents are bad at — sitting on a book for days and winning a seconds-long race at 11:40pm.

How SeatSwiper does it

SeatSwiper ships as an MCP connector. Connect it to Claude or ChatGPT and your assistant can set up and manage table watches conversationally: you describe the table, the agent arms a SeatSwiper watch, and SeatSwiper books the moment a matching table opens on Resy, SevenRooms, or OpenTable — a new date opening or a cancellation. Since the platforms themselves offer no official agent connector, this is currently the practical way to put “get me into Carbone” into an AI workflow at all.

The safety model matters and is worth insisting on from any tool in this space: the booking lands on your own reservation account, in your name, and you’re notified when it happens. Nothing is booked into a stranger’s name, resold, or transferred — that gray market is exactly what agent-driven booking shouldn’t automate. The agent’s authority ends at “set up the watch”; the reservation itself is indistinguishable from one you made by hand.

What to try

• Connect SeatSwiper as a connector/MCP server in your assistant, then ask it in plain language: “watch Tatiana for 2, any Thursday–Saturday in the next three weeks, after 8pm.”
• Let the agent manage the portfolio: it can list your active watches, adjust dates, or stand watches down as plans firm up.
• Pricing is the product’s standard deal: first booking free, then a flat $5 only on a successful booking, and a cancellation watch costs $1 to start (charged even if no table ever opens). The agent setting it up doesn’t change what you pay.