Browser Run: Giving Agents a Browser

Cloudflare has renamed Browser Rendering to Browser Run, and expanded its browser capabilities so AI agents can handle the web more reliably. The core idea is to offer the browser not as a simple rendering tool, but as an execution environment where agents can directly navigate, manipulate, observe, and hand off tasks.

The newly added features fall into five main areas.

• Live View: Check in real time what the agent is currently seeing and doing
• Human in the Loop: Let a human step in and take over at blocked points like logins
• Chrome DevTools Protocol (CDP) endpoint: Connect existing CDP automation scripts as-is
• MCP Client Support: Clients like Claude Desktop, Cursor, and OpenCode can use it as a remote browser
• WebMCP: Websites expose agent-facing tools directly on the page

The ways to control the browser have also broadened. In addition to the existing Puppeteer, Playwright, and Quick Actions, CDP is now exposed directly for low-level control. Existing CDP scripts written for self-hosted Chrome work almost as-is just by switching the WebSocket URL to Browser Run, and they can connect from any language or environment without a Cloudflare Worker.

WebMCP is an attempt to make the web more agent-friendly. Sites register tools via navigator.modelContext, let agents discover and execute them via navigator.modelContextTesting, and agents can then call explicit tools like search_flights for more reliable operation instead of guessing at the screen and clicking. This feature runs on Chromium 146+, and an experimental pool based on Chrome beta is also provided to try it out.

Observability and debugging have also been strengthened. Session Recordings log DOM changes, mouse and keyboard input, and page navigation so the cause of any issue can be traced, while Live View shows the current state including DOM, console, and network. At the same time, the concurrent browser limit has been raised from 30 to 120, and Quick Actions now handles up to 10 requests per second.

Additionally, the /crawl endpoint allows crawling an entire site with a single API call. Results can be returned as HTML, Markdown, or structured JSON, and you can configure crawl depth and scope, skip unchanged pages, and set include/exclude paths. However, Cloudflare has designed this not as a bypass-type crawler but as a 'well-behaved crawler' that follows robots.txt and AI Crawl Control, and it does not bypass bot defenses or CAPTCHAs.