Cloudflare wants to become the place where AI coding agents actually run, and it just added another big name to that pitch. The company announced support for running Cursor Cloud Agents inside Cloudflare Sandboxes, giving development teams a way to keep using Cursor while keeping the actual agent work confined to environments they control themselves.
The integration follows a pattern Cloudflare has already built with other agent platforms, including Devin Outposts and Claude Managed Agents. The underlying tension it’s solving is simple enough: developers want to stick with tools they already know, while the enterprises employing them want assurance over where that agent work actually executes, and what it can touch along the way.
Cursor Cloud Agents let developers hand off coding tasks from the Cursor app, the website, or Cursor’s mobile app. With Cursor Self-Hosted Machines, Cursor still handles the agent loop itself, inference, planning, orchestration, but the actual execution happens on a worker the customer chooses. Cloudflare Sandboxes now give that self-hosted worker a place to run inside the customer’s own Cloudflare account, rather than somewhere outside their control.
For developers, the workflow barely changes. They still start and manage agents through Cursor exactly as before, and Cursor still routes work and streams results back the same way. What actually shifts is location. Terminal commands, filesystem actions, and browser operations all execute inside sandbox environments the customer controls, which matters considerably for teams with strict rules about where code, build caches, and secrets can live.
Dane Knecht, Cloudflare’s chief technology officer, described the goal as giving developers tools that fit naturally into their existing workflow while giving enterprises a genuinely secure, isolated place to run that work. He framed the Cursor integration as another step toward Cloudflare positioning itself as the execution layer underneath the next wave of agentic applications, following similar moves with other AI platforms.
The connection model avoids opening any inbound access into a customer’s network. A Cursor worker instead opens a long-lived outbound HTTPS connection back to Cursor’s backend, and tool calls travel over that same connection. Enterprise teams can also set up worker pools, letting new agent sessions wait until capacity becomes available, then scale that capacity up or down automatically as demand shifts.
Repositories, build caches, and secrets stay on the customer’s own infrastructure throughout, while the system uploads only specific artifacts, screenshots, logs, and file references the model needs, so they can surface inside pull requests and dashboards.
