Browser Harness connects a large language model straight to a real Chrome instance over a single Chrome DevTools Protocol (CDP) websocket, so an agent works in the browser you already use and stay logged in to. When a task needs a capability the harness does not have, the agent writes the missing helper into its own workspace file and continues, so the helper set grows with each task.
Setup is done by a coding agent from a supplied prompt: it installs the package, registers the harness as a skill, and opens Chrome remote debugging so it can attach. The layout separates the parts an agent may change from the parts it may not — an install document handles the connection, a skill document teaches the browser workflow, and the package source stays protected while generated helpers live in a local agent workspace.
A companion MCP server, browser-harness-mcp, exposes the browser control helpers as Model Context Protocol tools over stdio, so any MCP client can drive the browser without a second CDP layer. For parallel work the project points at a hosted service that runs many browsers with live previews, proxies, stealth mode, and CAPTCHA solving; the local browser path is aimed at logged-in, personal tasks.
Features
- Direct CDP control: one editable websocket layer between the model and a real Chrome instance
- Self-written helpers: the agent adds missing helper functions to its workspace mid-task and reuses them later
- Coding-agent setup: a setup prompt installs the package, registers the skill, and connects the browser
- Protected core: the package source stays fixed while generated helpers stay in the agent workspace
- MCP server: browser control helpers are exposed as MCP tools over stdio for any MCP client
- Hosted scaling: an optional cloud service runs browsers in parallel with live previews, proxies, stealth, and CAPTCHA solving
