DoorDash wants AI agents to handle the office lunch run. On Wednesday, co-founder Andy Fang announced the DoorDash MCP for corporate ordering, which lets a team's agents place group orders, restock the office, schedule team lunches, track deliveries "and more."

We've seen a lot of enthusiasm around the DoorDash CLI, and people have been asking for it to work across their teams. Today, we're introducing the DoorDash MCP for corporate ordering, so your team’s agents can place group orders, restock the office, schedule team lunches, track… pic.twitter.com/A5EC4qqrKt

— Andy Fang (@andyfang) September 30, 2026

Quick explainer: MCP, short for Model Context Protocol, is a standard way for AI assistants to plug into outside apps and services. Instead of just suggesting what to eat, an agent connected to DoorDash's MCP can actually place the order.

According to Fang, this grew out of the DoorDash CLI, a command-line tool for developers. People liked it and kept asking for it to work across their teams.

Some companies are already on board. Fang lists SpaceXAI, Vercel, Cognition, Tempo and Mercor among the early adopters letting employees "order agentically." Early use cases include agents that restock office supplies and snacks based on inventory signals, and Slackbots that collect everyone's lunch requests and place the order.

DoorDash is also opening its dd-cli to more developers in the US and Canada, on macOS and Linux.

The company's AI research account followed up with more ideas for what an agent with DoorDash access could do: order items by voice through CarPlay so they arrive before you get home, turn a Slack thread into a single order for a 15-person team, find a dinner spot that fits your cravings and budget, or match meals to your macro and calorie targets using Apple Watch workout data.

Both the MCP and the CLI are in beta, and you can request access at developer.doordash.com/mcp. Neither post mentions pricing, and the US and Canada limit is only stated for the CLI.

Group lunch orders are a very real office headache, so this is a practical, low-stakes job for agents. We'd still want a human to glance at the cart before a 15-person order goes through, though.