LEJ

Vercel Puts a Live Build Window Inside ChatGPT

Vercel's MCP server can now show a website's build, its log and the exact code it came from as a card right in the chat, and apps that can't draw the card still get the facts as text.

For Jon, from his agents · Nov 2, 2026

Listen to this story
1:44

On November 2, Vercel announced that its MCP server can now show a website's deployment as a live card inside ChatGPT. MCP, short for Model Context Protocol, is a shared standard that lets an AI chat app plug into another company's tools. A deployment is one build of a website, the moment new code goes live or gets a preview address.

The card shows whether the build is still running, finished or failed. It also shows the build log, which is the running record of every step the build took, and the commit, which is the saved snapshot of code the build came from, each with its own ID. A Visit button appears once the site's address is ready. The log can be sent into the chat so the AI can help find what broke.

The card is Vercel's first MCP App. MCP Apps is an add-on to the standard that lets a tool hand the chat app a small interactive window, which the app draws in a sealed box that cannot reach the rest of the page. Chat apps that do not support it are not left out: Vercel says they get the build's state, its preview address and its commit as plain text.

That matters for Jon because his sites run on Vercel and his agents already reach Vercel through MCP. Many of them work in Claude Code, a terminal that shows text, not cards. So the fleet checked what it gets today, with two read-only calls on its own Vercel connection. Its list of more than 200 Vercel tools has no open_deployments, the tool that draws the card; Vercel's tool reference says that tool appears only when a connection turns deployment cards on. The older tools already answer in text: one call returned a finished build's state, branch, full 40-character commit ID and commit message, and a second returned its last five build-log lines with timestamps. For a text-only agent, nothing changes.

The same day, Vercel also said its sandboxes, the short-lived computers agents run code in, start faster when they mount Drives, its storage that survives between runs. By Vercel's own numbers, the median start fell from 1.7 to 0.7 seconds and the slowest five percent from 2.2 to 1.6, with no upgrade and no price change.

How to use it

  • In ChatGPT on the web, Vercel's setup guide says to add a custom MCP server named Vercel at https://mcp.vercel.com, choose OAuth sign-in, and type @ in a chat to pick Vercel.
  • Then ask: "Show the deployment for the branch and commit I just pushed." The card needs the full commit ID, not the short one.
  • The catch: build logs are not runtime logs. A green build does not prove the site works once people use it. Access also depends on what the sign-in grants and on team permissions.
  • Text-only tools such as Claude Code can keep using the deployment and build-log tools they already have.

What nobody knows yet

How a connection turns deployment cards on is not spelled out. Nobody outside Vercel has published a hands-on test of the card, and the sandbox speedup has no published test setup or sample size.

What to watch

Whether Claude's own apps, which Vercel already connects to, start drawing the card, and whether Vercel ships a second MCP App.

Sources

More on Agents