Cursor Lets Cloud Agents Run on Managed Machines
Cursor adds scalable worker pools across corporate infrastructure while retaining its cloud-based planning, inference and orchestration layer.
Execution moves inside the customer network
Cursor introduced Self-Hosted Machines that allow its cloud agents to execute commands on customer-managed laptops, servers and dynamically scaled worker pools. The agent interface, inference and planning remain in Cursor’s cloud, but repositories, build systems and command execution can sit inside an organization’s existing network.
A worker establishes an outbound HTTPS connection to Cursor, avoiding inbound firewall access. When an agent requests an action, Cursor sends the tool call to a dedicated worker, which edits files, runs commands and returns results. Cursor says it never initiates a connection into the customer network. However, tool outputs and transcripts still travel through its service for inference and may contain source code, so the architecture is not equivalent to a fully local agent.
Teams can attach an individual machine or create pools shared across repositories. A controller can start more workers when requests queue, release idle capacity and restore hibernated workspaces when follow-up instructions arrive. Cursor lists AWS Lambda, Coder, Cloudflare, Daytona, E2B, Modal, Namespace and Vercel among supported infrastructure and sandbox providers. Linux workers also gain browser control, joining existing Mac support and enabling agents to click, capture screenshots and test web applications.
The environment becomes part of the agent platform
Cursor says cloud agents now generate more than 60% of the pull requests merged internally. Giving comparable agents access to private services, bespoke build pipelines, GPUs and Apple hardware addresses a practical barrier to enterprise adoption: many valuable software tasks cannot be reproduced inside a generic hosted sandbox.
The move matters because coding-agent competition is expanding from model quality into workload placement, isolation and fleet operations. Cursor retains the high-value control plane while customers supply the machines and absorb more responsibility for credentials, capacity and containment. Enterprises considering the feature will therefore need to examine exactly what leaves each worker. Keeping execution inside the network reduces one class of exposure, but transmitted tool results mean it does not eliminate the data-governance boundary.