Skip to content
NewsProduct launch

Cursor ships self-hosted machine pools for cloud agents, opening the platform to security-constrained enterprise teams

· by Pondero Newsdesk

The short version

Cursor expanded its Self-Hosted Machines feature on September 2, 2026, adding dynamic worker pools, autoscaling, and Cloudflare Sandbox support so agent tool execution stays on customer-controlled infrastructure while the agent loop remains in Cursor's cloud.

Cursor ships self-hosted machine pools for cloud agents, opening the platform to security-constrained enterprise teams

Cursor shipped self-hosted machine support for Cloud Agents on September 2, 2026, per its changelog. Code, build outputs, and secrets stay on customer-operated infrastructure; the agent loop, which covers planning, inference, and orchestration, continues to run in Cursor's own cloud. That split is what enterprise security and compliance teams have been waiting for.

What the feature adds

The update introduced two execution modes. "My Machines" connects a single personal laptop or VM to a Cursor account for individual workflows. "Team Pools" create named queues of worker machines for teams or enterprises. Per the changelog, pools autoscale as requests arrive and shrink when workers disconnect, so capacity tracks demand without keeping idle machines warm continuously. Pools also support hibernation: an idle machine pauses, then restores within a reconnect window when the next task arrives.

Eight execution environments are supported at launch: AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, and E2B. Computer use on Linux and macOS is also available through self-hosted machines as of this release.

Cloudflare published a press release the same day confirming that Cursor Cloud Agents run natively on Cloudflare Sandboxes. Cloudflare Sandboxes are customer-account sandbox environments, separate from Cloudflare Workers. Terminal, filesystem, and browser tool calls execute inside those sandboxes while Cursor routes work and streams results back to the developer.

The architecture split

The design keeps a firm line between the two layers. Cursor retains the agent loop: it still handles task planning, model inference, and orchestration. The worker machine handles only tool execution, meaning filesystem reads and writes, shell commands, and browser interactions. Pools are not tied to any single repository, so a named pool can serve multiple projects within a team without reconfiguration.

From a compliance standpoint, sensitive code paths and secrets never leave the network the team controls, because the worker sits inside that perimeter. The Cloudflare integration extends this to teams that want execution inside their own Cloudflare account rather than an on-premises machine.

Why it matters

Security-conscious teams in financial services, healthcare, and regulated sectors have cited data residency as a direct blocker to adopting Cursor for production workflows. The self-hosted machine feature addresses that gap by letting a team bring its own execution environment without giving up the Cursor developer experience on the front end.

Cloudflare's simultaneous announcement signals a broader pattern. The press release mentioned prior integrations with Devin Outposts and Claude Managed Agents on Cloudflare Sandboxes. That positions Cloudflare Sandboxes as an enterprise execution layer for AI agent workloads across several platforms, not a Cursor-specific integration.

Teams on Cursor's Business or Enterprise plan can connect to the self-hosted pool system through the Cloud Agents dashboard. Try it at /go/cursor. Cursor has not yet published pricing for team pool capacity separately from the base seat price.

What to watch next

Whether GitHub Codespaces or Azure DevContainers announce Cursor integrations is the first signal to track. The second is separate pricing for team pool capacity, which would clarify how much of Cursor's enterprise deal value the self-hosted tier is expected to carry.

Sources