Cursor lets teams run cloud agents on machines they manage themselves

Cursor has launched Self-Hosted Machines, a feature that lets its cloud coding agents execute on infrastructure an organization controls, while Cursor's cloud continues to handle inference and planning for the agent.

What's new

As Cursor describes the split: "With Self-Hosted Machines, only the execution environment moves while the agent loop, inference, and planning remain in the Cursor cloud." A worker process, installed through the Cursor CLI (agent worker start), opens an outbound HTTPS connection from the customer's infrastructure to Cursor's cloud — Cursor never initiates an inbound connection into the customer's network. The agent harness plans and reasons in Cursor's cloud, then sends individual tool calls to the dedicated worker for execution, with results returned for the next inference cycle.

Workers can run individually as "My Machines" for a single developer, or be pooled for teams and enterprises, with Cursor automatically scaling the pool based on the queue of pending requests. Pools aren't tied to specific repositories: a request only needs to identify the target pool, and any available worker in it can claim the job — "Pools are not tied to individual repositories. A request only needs to identify the pool, and any available worker can claim it."

Context

Cursor positions the feature around a specific gap in its existing cloud-agent offering: agents that need direct access to internal services and source control, specialized hardware such as GPUs or Macs, or operating systems and build pipelines that are difficult to package into a standard Cloud Agent build image. Rather than requiring customers to adapt their infrastructure to Cursor's cloud build format, Self-Hosted Machines lets the execution environment stay wherever the customer's existing infrastructure already lives.

Why it matters

Cursor frames the need behind this release in terms of scale: cloud agents are increasingly driving enterprise software work, with the company saying they now generate more than 60% of internal pull requests at some organizations. At that volume, engineering teams need agents that can reach internal services, proprietary hardware, and non-standard build environments without Cursor operating that infrastructure itself — while keeping the planning and inference logic, which is where most of Cursor's product value sits, centralized in its own cloud. It's a pragmatic middle ground between a fully hosted agent product and a fully self-hosted one, aimed at enterprises with compliance or infrastructure constraints that a pure SaaS agent can't satisfy.