Vercel Opens Vercel Sandbox to Cursor Cloud Agents.

Cursor Cloud Agents can now execute inside Vercel Sandbox, giving development teams a serverless, isolated environment for testing agent-written code before it reaches production. The integration replaces Cursor’s hosted machines with a Vercel-managed execution layer while leaving Cursor responsible for the agent itself.

Vercel announced the integration on September 3, 2026. It targets a growing operational problem around autonomous coding agents: these systems need permission to clone repositories, modify files, install dependencies, execute commands, and run tests, but giving them access to persistent development infrastructure introduces security and maintenance concerns.

Vercel Sandbox handles that execution in a dedicated Firecracker microVM for each agent request. The result is an ephemeral workspace where an agent can work on code without sharing a long-lived machine with other sessions.

Cursor continues to manage its agent harness and inference loop, including the process that determines what the agent should do next. Its Self-Hosted Machines APIs then direct the work to an environment supplied by the customer. In this configuration, that environment is Vercel Sandbox rather than one of Cursor’s hosted machines.

The control plane combines Vercel Functions with Vercel Workflow. Together, they claim queued agent requests, provision Sandbox workers, monitor active sessions, retry failed work, and clean up resources when a task finishes.

That architecture provides four practical capabilities:

• Scale-to-zero workers: Teams do not need to keep virtual machines running while no agents are active.
• Per-request isolation: Every agent request receives its own Sandbox rather than sharing an execution environment.
• Durable retries: The control plane can retry work when a worker or session fails.
• Scoped credentials: Each Sandbox receives short-lived, user-specific credentials instead of persistent shared secrets.

The separation is important. Cursor remains the intelligence and orchestration layer for code generation, while Vercel supplies the disposable compute environment in which potentially consequential actions take place.

Coding agents are moving beyond code suggestions. A cloud agent may independently inspect a repository, edit several files, run a test suite, and prepare a change for review. That creates a larger security boundary than an autocomplete tool operating inside a developer’s local editor.

An isolated microVM limits the environment available to each request. Short-lived credentials also reduce the risk associated with leaving reusable secrets inside a worker. Automatic cleanup removes another source of operational overhead, particularly for teams that would otherwise maintain a fleet of self-hosted agent machines.

For small product teams, the scale-to-zero model may be equally significant. Agent capacity can be provisioned when a task arrives rather than reserved continuously. Vercel has not provided pricing or performance comparisons in the announcement, however, so teams will still need to evaluate the economics against Cursor-hosted machines or their own infrastructure.

The integration also fits Vercel’s broader effort to become infrastructure for AI agents, not only a destination for frontend deployments. That work includes controls around model usage, such as per-user budgets on AI Gateway, alongside execution, workflow, and credential-management products.

There is one notable access requirement. Cursor’s Self-Hosted Machines feature requires a Cursor Enterprise plan, according to the company’s Self-Hosted Machines documentation. Vercel’s changelog does not specify a separate Sandbox plan requirement or pricing for this integration.

Teams adopting the setup will therefore need both the relevant Cursor access and a configured Vercel control plane. Vercel points developers to a step-by-step guide from the announcement.

What is Vercel Sandbox?

Vercel Sandbox is an isolated execution environment that runs each agent request inside a dedicated Firecracker microVM. For Cursor Cloud Agents, it provides the workspace used to clone repositories, edit files, execute commands, and run tests.

Who can run Cursor Cloud Agents in Vercel Sandbox?

The integration is intended for teams using Cursor Cloud Agents with Cursor’s Self-Hosted Machines APIs. Self-Hosted Machines requires a Cursor Enterprise plan; Vercel did not state an additional Sandbox plan requirement in the announcement.

How does this differ from Cursor’s hosted machines?

Cursor still manages the agent harness and inference loop, but execution moves from Cursor’s hosted machines to Vercel Sandbox. Each request receives an isolated environment, while Vercel Functions and Vercel Workflow manage provisioning, monitoring, retries, and cleanup.