Give Claude Code, Codex or Cursor a real browser over MCP

What is a browser MCP server, and why a real browser?

A browser MCP server is a program that offers a coding agent a browser as a set of tools: open this address, read the page, take a screenshot, click that button, type here. MCP is the protocol those tools are described and called in, so one server works with every agent that speaks it. Most such servers start a fresh, headless browser that has none of your sign-ins.

Minda Super Intelligence ships the other kind: a developer bridge into its own built-in browser, the one you are signed in to, with each workspace’s accounts kept apart. A coding agent can open your staging admin behind its login, your CRM, your shop’s back office, the page you are debugging in the state you see it, instead of a blank browser that has to log in first or cannot. It is off by default, on this computer only, and on every plan.

What can the coding agent do through it?

The tools are the SI agent’s own, so a coding agent gets what the chat agent gets, by name:

• tabs lists the open tabs; open loads an address or search words in a new background tab; navigate loads an address in a tab the task owns.
• snapshot lists what can be clicked or filled, numbered; read gives the page’s text, or with a question the passages that answer it; screenshot a picture of the visible part; extract the page’s main table or list as rows.
• click and type by the number or label from the snapshot; press_key and scroll.
• finish ends the task and closes the tabs it opened, unless keep is set.
• Minda’s live list may add more, such as read_pages, web_search, fill_form, wait_for, go_back, close_tab and search_knowledge; a tool another module adds is offered as soon as it exists. Run node minda-mcp.mjs tools to see the list on your computer.

How do I turn the bridge on?

1. In Minda, open Settings → Developer and turn on Let coding agents control Minda. The status line says Listening.
2. Leave Ask before actions on while you learn it: an agent then waits for your OK in Minda before it clicks, types or presses keys on a page. Using one of your own tabs, and searching your history, always ask.
3. Nothing else to configure: Minda listens on 127.0.0.1 on a random port and writes a connection file, automation.json, into your profile folder with the port and a token, readable by your user only. The token is the only credential, and New token stops the old one at once.

How do I connect Claude Code, Codex, Cursor or OpenClaw?

Once per agent. Settings → Developer shows each snippet with the real path on your computer; the shapes are these. The agent then starts Minda’s MCP server by itself, which needs Node.js 18 or later and reads the token from the connection file on every call, so a new token or a restart of Minda needs no restart of the agent.

What is a good first task?

1. Open the page you want tested in Minda and sign in, in the workspace you use for it.
2. Tell your coding agent, in its own chat: "Use the minda tools. Open the orders page of our staging admin, snapshot it, click New order, type Test in the customer field, take a screenshot and finish." Name a task so its tabs stay together.
3. Watch Minda: a card asks for your OK before the first click while Ask before actions is on. The agent’s tabs open in the background in the active workspace; the tab you are looking at stays on screen.
4. Read the agent’s report, then check Settings → Developer → Recent calls: the tool, the task, the site and the result of each call, never page text or what was typed.
5. Once it behaves, turn Ask before actions off for the page actions you trust; a tab the task did not open, and your history, still ask.

What keeps it safe?

• Off by default, and a team policy (DeveloperAutomationEnabled) can keep it off on managed computers.
• This computer only. It listens on 127.0.0.1 on a random port; a request with an Origin header (a web page) or with another Host than 127.0.0.1 and that port (DNS rebinding) is refused.
• One secret. Every request needs the 32-byte bearer token from the connection file, compared in constant time; the file is readable by your user only, rewritten with a new token at each start and deleted on stop and quit.
• Never a private window. Private windows are never reachable; a task owns the tabs it opens, in the window’s active workspace, driven off screen, and finish closes them.
• Your OK where it counts. Each page action waits for your answer in Minda while Ask before actions is on (two minutes, then denied); a tab the task did not open, and your history, always ask.
• A log, not a transcript. Recent calls keeps the tool, the task, the site and the result, never page content or typed text.

How does it compare with a headless browser?

A headless browser is the right tool for a test suite; the bridge is the right tool for the page behind your login, on your machine, with you watching.