Let your AI agent use your browser -- the real one, with your logins, your tabs, your sessions -- from any MCP client. Governed when you want it, wide open when you don't.
A thin extension drives the Chrome or Edge you already use. Real cookies, real SSO, real tabs -- never a fresh profile, a profile copy, or someone's cloud browser.
Claude Code, Cursor, VS Code, Zed, Cline -- anything that speaks MCP. One portable Rust binary, zero runtime dependencies, works with Claude on Bedrock or Vertex too.
Capability grants (read / action / write / execute), never-touch domains, a pause button and a panic switch, and a structured audit trail. All-open is the first-class default; policy is opt-in.
Every action the agent takes is visible in the browser: click ripples, a comet trail on drags, a shimmer while it types, captions narrating each step. You always know what your agent is doing, and you can grab the wheel at any moment.
The tool surface preserves the schemas Claude was trained on, byte for byte, and adds composition tools on top: multi-step scripts with dry-run, semantic form filling, and page-state awareness.
If you use Claude Code on a direct Anthropic plan, it ships its own Chrome
integration and it is good -- claude --chrome may be all you need. Ghostlight is
for everyone that path excludes: other MCP clients, Claude via cloud providers, and anyone
who wants policy as code, an audit trail, or a self-hosted stack. The
honest
comparison is part of the docs.
curl -fsSL https://raw.githubusercontent.com/sylin-org/ghostlight/main/scripts/get.sh | sh
Windows: irm https://raw.githubusercontent.com/sylin-org/ghostlight/main/scripts/get.ps1 | iex
-- then add the browser extension. The install page has
one-click buttons for Cursor and VS Code, the npx path, and the manual route.