Glasspane — Give any AI agent a real browser

How it works

Three steps from zero to agent-controlled browser.

Deploy

Run docker compose up to launch Chrome and the MCP server in one command.

Connect

Point any MCP-compatible agent at the server URL. It discovers tools automatically.

Observe

Watch the agent work in real time through the co-browse UI. Debug, intervene, take over.

Built for agents, visible to humans

Everything an AI agent needs to work with the real web.

MCP Tools

Navigate, click, type, screenshot, extract DOM — a full browser automation toolkit exposed as MCP tools any agent can call.

Co-browse UI

Watch the agent navigate in real time through a live-streaming browser view. Step in and take manual control at any point.

Plugin System

Extend Glasspane with custom MCP tools. Add a plugin folder, export a register function, and allow it explicitly with PLUGINS_ALLOW.

Auth Persistence

Chrome runs persistently with a real user profile. Log in once — cookies, sessions, and credentials survive across agent runs.

See it in action

An agent navigating, extracting data, and completing tasks — all visible through the co-browse UI.

Demo video coming soon

Up and running in 60 seconds

One clone, one command. No build step required.

bash

# Clone the repo
git clone https://github.com/anthropics/glasspane.git
cd glasspane

# Configure
cp .env.example .env

# Launch
docker compose up -d

# Open the co-browse UI
open http://localhost:7600

Then point your MCP-compatible agent at http://localhost:7600/mcp and it will discover all available tools.