agent-browser

agent-browser gives AI agents a compact text view of a browser page, with references that identify the exact elements they can interact with. It's for developers who want coding assistants or other agents to use websites without filling their context with a full page's HTML. The open source tool runs locally on macOS, Linux and Windows under the Apache 2.0 license.

Agents can follow links, fill forms, manage tabs and interact with local PDFs or HTML files. Screenshots, video recording and streaming let you inspect what the browser does, while network controls, profiling and page comparisons support testing and debugging. Sessions and saved authentication state help agents continue work across browser interactions.

It works with Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Google Gemini and opencode, as well as any agent that can execute shell commands. Its Rust runtime controls Chrome directly without requiring Node.js or Playwright for the daemon. Chrome is the default browser; it also supports Lightpanda and Mobile Safari in the iOS Simulator, which requires macOS and Xcode. Browser provider plugins can connect it to cloud browsers instead of a local browser.

Optional security controls include an encrypted local credential vault that keeps passwords out of the LLM's context, domain restrictions and policies that gate destructive actions. Page output can carry boundary markers to help agents distinguish website content from tool output. Plugins can add credential providers, browser providers and domain-specific commands.