Building a Cloud Environment for Coding Agents

Cursor redesigned its development environment like a product so that cloud agents could run and test changes on its own codebase. As a result, the share of merged PRs created by cloud agents in the Cursor monorepo has grown from about 10% last December to over half now.

First, they narrowed the gap between local development environments and cloud VMs. Unlike the Mac environment most developers use, cloud VMs are Linux-based, so they modified development utilities and setup scripts to work on Ubuntu and included core dependencies in a Cursor-specific Dockerfile.

On the security side, they added the following features to safely inject secrets needed by the agent environment.

• Network egress restrictions
• Scoped and proxied Git remote repository access
• Secret scanning of commits and commit messages
• Secret masking that removes secret values from tool results

Even after setting up the environment, agents still had to deal with complex build commands, flags, and utility scripts, which prevented them from running code reliably. To solve this, they created a CLI called anydev that starts all services and consolidates key utility scripts, providing multi-level --help guidance for each subcommand.

anydev also includes a supervisor process that watches long-running build commands and automatically restarts them. Since agents no longer need to memorize complex commands or manage processes directly, the reliability of code execution and testing improved significantly.

Once combined with computer use features and the recordScreen tool, cloud agents became able to test changes end-to-end and demonstrate the results to users. Engineers can now receive demo recordings made by agents via Slack or PRs, and even merge and deploy code without checking out the branch locally.

To address the problem of the codebase and surrounding environment continuously changing, a system for diagnosing and recovering the environment was also needed. To this end, Cursor built Cursor Cloud MCP, which can dynamically discover tools, laying the groundwork for identifying and recovering from abnormal environment states.