browser-agent
Drives the user's real Brave browser (a Chromium 116+ browser) through an extension that connects out to the server over WebSocket, letting agents open, group, read, click, fill, upload, screenshot, and close tabs and run JavaScript in the user's logged-in sessions.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
browser-agent
An MCP server that lets AI agents drive your real browser through a Chrome extension.
agent ──MCP (HTTP)──▶ browser-agent server ◀──WebSocket── Chrome extension ──▶ your tabsThe extension connects out to the server, so the server can run anywhere: your laptop, a VPS, or a PaaS.
Agents keep their tabs in tab groups named for their task:
browser_navigatetakes agroupname for a new tab, and tabs a page opens join its group.Agents read pages as a compact outline with
[eN]refs. After every action they get a diff of what changed.One browser at a time. If a second extension connects, it replaces the first.
Forked from qwadrox/bb-plugin-desktop-browser-agent. The CUA driver, the bb plugin parts, native messaging and most of the guardrails were removed.
No guardrails. Anyone who has the MCP URL can do anything in your browser that you can, including run JavaScript in your logged-in sessions. Keep the
SECRETsecret.
Tools
Tool | What it does |
| List the open tabs with ids, titles, URLs and groups |
| Open a URL. With no |
| Read the page as an outline with refs; |
| Click a ref or an x/y point with trusted mouse input: right/middle button, double/triple click, modifiers, hover, and drag (including HTML drag and drop) |
| Give files to a file input, or drop them on an element, from a path on the browser's computer, a URL, or base64 data |
| Console messages, page errors and network requests, recorded from the first call on a tab; |
| Fill text fields, selects and checkboxes in one call |
| Press a key or a combo ( |
| Run an async JS function body, in an isolated world (default) or the page's |
| Take a JPEG of the tab or of one ref |
| Close a tab |
If the page re-rendered an element, so its ref is gone, actions use the one element with the same role and name and say so.
Actions take an optional wait_for text to wait for after the action. Dialogs (alert, confirm, prompt) are accepted automatically and reported in the result.
Related MCP server: tablaze
Run the server
Requires Bun.
bun install
SECRET=$(openssl rand -hex 16) PORT=3000 bun startEnv | Default | |
|
| |
| (none) | Path secret. The endpoints become |
Endpoints:
POST /mcp/<SECRET>?agent=<name>: stateless Streamable HTTP MCP.agent(or anX-Agent-Idheader) is optional: it's the group name when the agent doesn't passgroup. The default isAgent.GET /ext/<SECRET>: the extension's WebSocket.GET /healthz:{"ok":true,"browser":{…} | null}.
To run it with Docker:
docker build -t browser-agent .
docker run -p 3000:3000 -e SECRET=... browser-agentBehind a reverse proxy, make sure the proxy passes WebSocket upgrades through on /ext/…. Tool calls can take up to 90 s, so allow long requests.
Install the extension
bun run build:extensionIn Chrome (116+) or Brave, open chrome://extensions (brave://extensions), turn on Developer mode, click Load unpacked and pick extension/dist.
Click the extension's icon, paste the server's extension URL, https://<host>/ext/<SECRET> (or wss://…), and press Save. The popup shows whether it is connected. It reconnects on its own.
To bake a default URL into a build instead (the popup still overrides it): SERVER_URL=wss://<host>/ext/<SECRET> bun run build:extension.
While it is driving a tab, Chrome shows a "started debugging this browser" bar.
Connect an agent
Claude Code:
claude mcp add --transport http browser "https://browser.example.com/mcp/<SECRET>?agent=claude"Any MCP client that speaks Streamable HTTP:
{
"mcpServers": {
"browser": { "type": "http", "url": "https://browser.example.com/mcp/<SECRET>?agent=research" }
}
}Agents name their own groups per task. agent= only sets the fallback name. Agents can still see and use every tab.
Develop
bun run dev # server with --watch
bun run typecheck
CHROMIUM="<Chrome for Testing binary>" bun run test:e2eThe end-to-end test starts the server, builds the extension against it, loads the extension into headless Chrome for Testing, and drives it over MCP. Branded Chrome ignores --load-extension, so get Chrome for Testing with npx @puppeteer/browsers install chrome@stable.
Layout:
server/src: Bun server (server.ts), extension bridge (bridge.ts), MCP tool definitions (tools.ts)extension/src: service worker (background.ts), popup (popup.ts), in-page functions (page.ts,outline.ts), wire protocol (shared.ts)
License
MIT. See LICENSE.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
Give AI agents a real cloud Chrome: open sessions, navigate, read pages, click, type and screenshot.
Hosted real Google Chrome MCP with per-user persistent state. Navigate, click, type, screenshot.
Web scraping, browser automation, screenshots and AI data extraction for AI agents via MCP.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables AI coding agents to control and inspect a live Chrome browser via MCP, including performance tracing, network and console debugging, screenshots, and reliable Puppeteer-based automation.2,562,190 npmApache 2.0
- AlicenseAqualityBmaintenanceEnables AI agents to drive a Chromium browser through MCP, with persistent sessions, compact page snapshots, guarded actions, batched operations, and explicit result verification.16MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to see and control the user's real Chrome/Brave/Edge profile over MCP, so they can read pages, click, type, take screenshots, audit layouts, debug CSS, and scrape paginated or infinite-scroll data. Because it drives the normal browser via the DevTools protocol with real input events, it works on modern JavaScript apps and on sites where the user is logged in.17 npm1MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to drive a live, persistent self-hosted Chromium session over MCP, running browser actions via Playwright/CDP, capturing page snapshots, screenshots and videos, and managing uploads and downloads in the same tabs and signed-in accounts a human user can take over at any time.28 npmApache 2.0