Connect a cloud AI agent to a browser on your own computer

A cloud AI agent can drive a browser on your own computer through a remote MCP server that relays its tool calls. The agent connects to the server over the internet and signs in; the browser app on your computer keeps one outbound connection to the same server; each call, such as "open this page" or "click that button", is passed down to your computer and run in a real browser profile there. Nothing on your computer is opened to the internet.

In Browz that relay is the hosted MCP server at mcp.gobrowz.com, a free public beta for every account, Free included. This post explains why cloud agents need it, how it works, what it offers and keeps local on purpose, and how to connect Claude and OpenAI agents.

Why a cloud agent needs a browser somewhere

Agents increasingly run in the cloud: a hosted chat app with connectors, an agent built on a model provider's API, a scheduled job on a server. They are good at reasoning about pages, but they have no browser of their own on your side. When a task needs a real website, the browser has to live somewhere.

There are two broad answers:

	A browser in a cloud service	A browser on your own computer, through a relay
Where the browser runs	A data centre	Your computer
Network	The service's addresses	Your own connection, or the proxy you give the profile
Profiles and sessions	Created in the service	Your own Browz profiles, with the cookies they already hold
Fingerprint	Whatever the service's browsers present	Each profile's own consistent device
A person nearby	Rarely	You, at the same computer, for hand-off

Cloud browser services suit large, anonymous workloads. A relay to your own computer suits work that belongs to you: accounts you already use, profiles you have set up with their own proxies, and a person who can step in for a sign-in.

How the Browz relay works

Remote control on. In the Browz app, turn on Settings, Account, Remote control. The app opens one outbound connection to our server and keeps it. Remote control uses your computer's hardware security chip: the Secure Enclave on a Mac, or TPM 2.0 on Windows and Linux.
The AI app connects. You add https://mcp.gobrowz.com/mcp to the AI app as a remote MCP server or custom connector. It speaks Streamable HTTP.
You sign in and consent. The AI app opens a Browz sign-in page. You sign in, see what the app is asking for, choose the workspace and allow it. A client that cannot sign in with a browser uses an agent token instead, kept in an environment variable on the machine that runs it.
Calls flow down. Each tool call goes from the AI app to the server and down the app's connection to your computer, where Browz runs it in the profile and sends back the result.

If more than one of your computers is online with Remote control, the agent calls browz_devices_list and passes a deviceId to choose one.

The 17 tools, and what stays local on purpose

The hosted server offers 17 tools: your computers, profiles (list, create, clone, start, stop, delete), tabs, pages (navigate, snapshot, act, wait and screenshot), hand-off to a person, and ending a session. The tool names and inputs match the local server's, so a prompt written for one works with the other.

On purpose, it offers no cookies, no running of script, no downloads and no stored sign-ins. Those stay with the local MCP server on your computer. An agent in the cloud reads and acts on pages, and hands sign-ins to you.

Hand-off from the cloud

When a page needs a person, the agent calls browz_request_human. The profile's window comes to the front on your computer and you finish the sign-in, code or CAPTCHA. Over the hosted server one call waits about 70 seconds and returns state: "waiting" with a handoffId; the agent calls again with it to keep waiting. Human hand-off for AI agents covers the pattern.

The free public beta

While the hosted MCP is in beta, every account, Free included, gets the free beta allowance for hosted agents:

2 hosted sessions at once;
2,000 actions and 300 relay minutes a month, reset at 00:00 UTC on the 1st;
30 tool calls a minute and 2,000 in any 24 hours for each connected agent.

A paid plan keeps its own hosted allowance, and Pro 50 and above include more. Agents on your own computer, through the local MCP server, are never counted. When an agent reaches a limit, the tool result says which one and when it can try again. The current figures are on hosted agents.

What passes through our server

Page snapshots, screenshots and the text an agent types pass through our server encrypted. They become unreadable after 7 days and are then deleted. The action log (the tool, the profile, the page's origin and path, the outcome and the time, without page content) is kept for 90 days. Our privacy policy has the rest.

You stay in control

On your account page, under AI agents, you see every connected AI app and agent token with what it did. You can limit one to some profiles or to read-only, limit the sites it may open, stop it now, or revoke it. Turning off Remote control in the app cuts every hosted agent off at once.

Set it up

Install Browz from the download page, sign in, and turn on Remote control first.

Claude.ai

In Claude.ai, go to Customize, Connectors, Add, Add custom connector. Name it Browz, paste https://mcp.gobrowz.com/mcp, choose "Sign in now" and sign in with your Browz account. In a chat, turn Browz on under the + menu's Connectors. Which Claude plans offer custom connectors is in Anthropic's own documentation.

Claude Code, from a remote machine

bash

claude mcp add --transport http browz https://mcp.gobrowz.com/mcp
# then run /mcp in Claude Code to sign in with your Browz account

OpenAI Responses API

Create an agent token in the app under Automate, Connect an AI agent, or on your account page, and keep it in BROWZ_AGENT_TOKEN:

import OpenAI from 'openai';
const openai = new OpenAI();
const response = await openai.responses.create({
  model: 'gpt-5',
  tools: [{
    type: 'mcp',
    server_label: 'browz',
    server_url: 'https://mcp.gobrowz.com/mcp',
    authorization: process.env.BROWZ_AGENT_TOKEN,
    require_approval: 'never',
  }],
  input: '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.',
});
console.log(response.output_text);

OpenAI Agents SDK

In Python, HostedMCPTool with the same server URL and token lets OpenAI call Browz for the agent. If the agent runs in your own process, MCPServerStreamableHttp connects to the server directly with Authorization: Bearer and the token; the OpenAI Agents recipe is a complete project for that. More clients are in the setup docs and on Browz for OpenAI agents.

Check it works

Ask the agent: "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." It should answer "Example Domain", and you will see the profile open and close on your computer.

Why run it this way

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. The relay keeps that promise for cloud agents: the browser and its profiles stay on your computer, each profile keeps its engine-level fingerprint and its own proxy, sensitive tools stay local, and you can cut every hosted agent off with one switch. Browz for Claude has the Claude setups in one place.

Frequently asked questions

What is a remote MCP server?

It is an MCP server that an AI app reaches over the network, usually with Streamable HTTP and an OAuth sign-in, rather than starting it on the same computer. Cloud-hosted AI apps can only use remote servers.

Can Claude.ai control a browser on my computer?

Yes, through a custom connector to a remote MCP server. With Browz, you add mcp.gobrowz.com as a connector, sign in with your Browz account, and Claude.ai drives Browz profiles on your computer while the app is open with Remote control on.

Is the Browz hosted MCP server free?

Yes. It is a free public beta for every Browz account, Free included, within the beta allowance: 2 hosted sessions at once, 2,000 actions and 300 relay minutes a month. Agents that use the local MCP server on your own computer are not counted at all.

Does the hosted server open a port on my computer?

No. The Browz app keeps one outbound connection to the server, and calls come back down that connection. Nothing on your computer listens for connections from the internet.

Why does the hosted server leave out cookies and stored sign-ins?

By design, the most sensitive tools stay on your own computer. A cloud agent reads and acts on pages and hands sign-ins to you, while a local agent with the right key can use cookies, script, downloads and stored sign-ins.