AI agents / MCP browser
You decide exactly what each agent may do: its own key, its own profiles, only the sites you allow, sign-ins it can use on your computer but never read, and a person for the login. Browz ships with an MCP server: any MCP client can open an isolated browser profile, read the page as a compact snapshot, click and type by reference and take screenshots, all in a real browser on your computer with its own fingerprint, proxy and cookies. Run it over stdio with npx -y @browz/mcp, or over HTTP on 127.0.0.1. Underneath, every profile’s fingerprint is set in the browser engine itself, not by a script on top, and you can test it yourself (engine-level vs injected scripts).
claude mcp add --transport stdio browz -- /Applications/Browz.app/Contents/Resources/browz-mcpBrowser MCP server
Each agent gets its own key, limited to the profiles and sites you choose, with call limits, a log on your computer, revoke and “stop all agents”. On your computer it can sign in with a password and code you stored without ever seeing them, and it hands logins, 2FA and CAPTCHAs to a person.
Set an agent’s limitsbrowz_snapshot returns the page as an outline of roles, names and states, sized for the model, and each interactive element carries a short ref. browz_act clicks, fills, types, selects or scrolls by ref, with real input events and no script injected into the page.
browz_session_open gives a task its own profile: throwaway, or returning for the same identity. browz_session_resume brings it back after a crash; browz_session_close wipes a throwaway one.
The client starts the server inside the app over stdio, with nothing to paste, or connects to it over HTTP on 127.0.0.1 with a bearer file. Hosted agents use https://mcp.gobrowz.com/mcp, in beta: they sign in with OAuth and get only the access you allow.
Setup
These are the same snippets as the setup docs and the Connect an AI agent screen in the app.
The server is in the app: browz-mcp. Download Browz.
Claude Code takes one command; most other clients take the same JSON entry under mcpServers.
Ask: “Using Browz: create a new profile called agent-check, start it, open https://example.com, tell me the page title, take a screenshot, then stop the session and delete the profile.” The answer should be “Example Domain”, with no profile left behind.
Install Browz, open it, and your agent can start: no account needed for local use. Keep the app open while your agent works; it can stay on its sign-in screen.
Claude Code Add Browz to Claude Code
claude mcp add --transport stdio browz -- /Applications/Browz.app/Contents/Resources/browz-mcpAny client that reads mcpServers Add this entry to your client’s MCP config file (for Claude Desktop, claude_desktop_config.json; for Cursor, mcp.json), then restart the client.
{
"mcpServers": {
"browz": {
"command": "/Applications/Browz.app/Contents/Resources/browz-mcp",
"args": []
}
}
}Remote (Streamable HTTP, OAuth), beta Add the remote server; Claude Code signs you in with your Browz account (run /mcp)
claude mcp add --transport http browz https://mcp.gobrowz.com/mcpThe snippets use the macOS path to the server inside Browz, /Applications/Browz.app/Contents/Resources/browz-mcp; on Windows and Linux the paths are in the setup docs. Give each agent its own key, limited to the profiles and sites it needs, and let it sign in without seeing the password. The remote MCP server for hosted agents is in Beta: free for every account, Free included, within the beta allowance, and it needs the Browz app open with Remote control on, on a computer with a hardware security chip. See what it needs and its allowance. Use agents only where each site’s terms allow it: our acceptable use policy applies to agents the same as to people.
Tools
The tools of the MCP server in the current Browz app. The remote server, in beta, offers the profile, tab and page tools; cookies, JavaScript, downloads and stored sign-ins stay local.
browz_session_open
browz_session_resume
browz_session_close
browz_session_list
browz_profile_list
browz_profile_create
browz_profile_clone
browz_profile_start
browz_profile_stop
browz_profile_delete
browz_tabs_list
browz_tab_open
browz_tab_close
browz_navigate
browz_snapshot
browz_act
browz_wait
browz_screenshot
browz_session_end
browz_cookies_get
browz_cookies_set
browz_evaluate
browz_download_list
browz_download_read
browz_running
browz_group_list
browz_proxy_list
browz_request_human
browz_credential_list
browz_credential_fillbrowz_session_open → a fresh, isolated profile
browz_navigate → https://example.com
browz_snapshot → the page, with refs like e12
browz_act → click e12
browz_wait → text “Results”
browz_screenshot → PNG for the model
browz_session_close → profile wipedFAQ
A browser an AI agent controls through the Model Context Protocol: the agent calls tools such as navigate, snapshot, click and screenshot, and the browser carries them out. Browz is one where every task can get its own isolated profile with its own fingerprint, proxy and cookies.
On your computer, stdio (the client starts the server inside the app) or Streamable HTTP on 127.0.0.1 with a bearer file. Remotely, Streamable HTTP at https://mcp.gobrowz.com/mcp with OAuth sign-in, in free public beta.
Either. Profiles open as ordinary windows you can watch and take over, or headless when the agent asks for it.
More for agent builders
Everything an agent gets in Browz.
OverviewOne consistent device, its own proxy and cookies for every agent task.
Read moreClaude Desktop, Claude Code and Claude.ai, set up in a minute.
Read moreAn isolated browser for Cursor’s agent, from one config entry.
Read more