Cursor Cloud Agents now run on your own infra via Vercel Sandbox

Firecracker microVMs and scale-to-zero pools move agent execution inside your own network — but the self-hosted APIs are Cursor Enterprise-only.

Nowline SEP 4 4:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Agents execute in a microVM you control

    Cursor Cloud Agents can now run inside Vercel Sandbox: Cursor keeps the agent harness and inference loop, while Vercel spins up a dedicated Firecracker microVM per request to clone your repo, edit files, and run tests. Your code and short-lived, user-scoped credentials never leave your Vercel account.

  • Eight backends, one outbound tunnel

    Cursor's self-hosted machines let those agents run on AWS Lambda, Cloudflare, Coder, Daytona, E2B, Modal, Namespace, or Vercel. Workers hold a long-lived outbound HTTPS connection to Cursor's cloud, so Cursor never dials into your network — the worker always reaches out.

  • The catch: Enterprise-only

    The Self-Hosted Machines APIs require a Cursor Enterprise plan plus an admin-issued team service-account key. Solo builders and Pro teams can't flip this on yet — it's aimed at orgs that need agents sitting next to private source control and internal services.

  • Scale-to-zero, with the timeouts to know

    Pools register with zero workers and only provision a Sandbox when a request lands, then tear it down. Defaults worth memorizing: agent status is polled every 30s, a worker exits after a 10-minute idle grace, and each Sandbox hard-caps at a 45-minute timeout.

  • Build it this weekend

    The reference implementation is a Next.js app: Vercel Functions expose the controller endpoint and Vercel Workflow handles durable, deduplicated orchestration; you bake a worker snapshot with the Cursor Agent CLI pre-installed. Prereqs are Node 22+, the Vercel CLI, and Sandbox access.

  • Elsewhere: cheaper Vercel builds

    Separately, Vercel added Basic build machines (2 vCPU, 8 GB) as a lower-cost tier for Pro and Enterprise — pick it when your builds don't need the default machine's headroom and you'd rather not pay for it.