RELATED 01
Browser Use
OFFICIAL SERVERLocal MCP server that drives a Chrome or Chromium browser step by step, with an agent fallback
120kGitHub stars
8.1M/moPyPI downloads
119popularity score
BrowserOS neo
A separate desktop Chromium browser that AI agents drive through a local MCP endpoint, signed in as you
Official · BrowserOS TypeScript Local · Streamable HTTP Browser automation
89 popularity score ★ 14k ↓ 482/wk
Connect your client
Local (once it is running)
Add in one click
Opens the app, which asks before it adds the server.
Or add it by command or config
Claude CodeTerminal
claude mcp add --transport http browseros-neo http://127.0.0.1:9200/mcpCursor~/.cursor/mcp.json
{
"mcpServers": {
"browseros-neo": {
"url": "http://127.0.0.1:9200/mcp"
}
}
}VS Code.vscode/mcp.json
{
"servers": {
"browseros-neo": {
"type": "http",
"url": "http://127.0.0.1:9200/mcp"
}
}
}Codex~/.codex/config.toml
[mcp_servers.browseros-neo]
url = "http://127.0.0.1:9200/mcp"Gemini CLITerminal
gemini mcp add --scope user --transport http browseros-neo http://127.0.0.1:9200/mcpWhat it does
BrowserOS neo is a desktop browser built on Chromium for AI agents to use, kept apart from your everyday Chrome. Once it is installed and running, it serves an MCP endpoint on your own computer (http://127.0.0.1:9200/mcp by default). Claude Code, Codex, Cursor, VS Code, Zed, OpenCode, Antigravity and other MCP clients connect to that address and control the browser through it.
The main difference from headless tools such as Playwright is that the agent works inside a browser that is already logged in. One click copies logins, bookmarks and extensions over from Chrome, and they stay between sessions. Each agent gets its own tab, so several tasks can run in parallel. A new-tab dashboard shows what each agent is doing, and every session is saved to disk as a video with a step-by-step timeline you can replay.
The tools work page by page. An agent takes an accessibility-tree snapshot, acts on elements by reference, and gets back a diff of what changed. It can also read pages as markdown, search them, take screenshots and PDFs, handle downloads and uploads, and manage tabs, tab groups, windows and history. A run tool executes a short script in a sandbox to string many steps together. When the agent hits something only a person can do, such as a captcha, a one-time code or a sign-in, it can ask the user for help.
The same repository also ships the original BrowserOS, a Chromium fork for people that has its own built-in agent and its own MCP endpoint, with a larger set of tools. This page covers neo, which is the product the README leads with.
What it needs
You need the BrowserOS neo desktop app installed and running: Homebrew cask browseros-neo or a direct download for macOS, Windows, or Linux x64 (AppImage or DEB). You also need an MCP client that can connect to a streamable HTTP URL. Clients on the one-click list are set up from the app's MCP page. Claude Desktop uses a one-file extension. Any other client takes the endpoint URL by hand.
Limits
The endpoint only works while the app is running and listens only on 127.0.0.1, so remote agents need a tunnel. One run script is capped at 30 seconds of wall-clock time. The skill file advises about five new page loads per run and no more than five agent tabs at once unless the user asks for more. Element references go stale when the page changes, so the agent has to take a new snapshot. It does not replace Chrome. It is a second browser that runs alongside it.
Who it suits
BrowserOS neo suits people who want their coding agent to do web tasks that need their real accounts, such as email, social posts, CRMs or internal tools, from their own machine and IP address, with a replay of each session. If you need throwaway headless browsers for CI or test runs, a Playwright-style driver fits better; the README itself says neo is not meant for that. It also does not suit agents on a server with no desktop, because the browser has to run as a desktop app.
Tools 20
tabs- List, open, switch between and close browser tabs
tab_groups- Create and manage groups of tabs
history- Look up and manage the browsing history
navigate- Load a URL in a tab, or go back, forward or reload
snapshot- Return the page as an accessibility tree with references to its elements
diff- Show what changed on the page since the last look
act- Click, type or fill elements by reference, including whole forms at once
download- Save files linked from a page
upload- Attach local files to a page's file inputs
read- Extract page content as markdown
grep- Search a page's text without returning the whole page
screenshot- Capture an image of the page
pdf- Save the page as a PDF
wait- Wait until a given text or selector appears
windows- Open, close and arrange browser windows
evaluate- Run JavaScript inside the page
run- Run a sandboxed script that chains many browser steps in one call, up to 30 seconds
name_session- Label the agent's session with a name, category and summary for the dashboard
request_human_help- Ask the user to step in for a sign-in, code, captcha or approval
await_human_help- Keep waiting for the user to finish a requested hand-off