AI agents

An AI agent such as Claude Code, Cursor or Codex can work in your Nevazu browser: it takes screenshots of the whole remote screen, clicks, types, scrolls, presses shortcuts and uploads files, in the same Chrome with the same sign-ins you use in the live view. You watch in the live view at the same time, and your own click takes control back at any moment.

Turning it on or off

Allow connections is ticked when you create a browser; untick it to create a browser no AI agent or chat app can ever reach. In a browser’s Settings → Agent access the same switch can be changed later: turning it off disconnects all AI agents and chats at once and keeps their sign-ins for when you turn it on again.

Connecting an agent

Nothing needs to be installed, and your agent sets itself up. Each browser has its own server address for the Model Context Protocol (MCP), which AI agents use to reach tools. The agent connects to it over HTTPS and signs in with your Nevazu account, which you approve in your web browser; there is no key to store.

On the browsers page or the Agents page, choose the robot button on a browser, or Connect an agent. The Agent Connection panel has three sections: Use in chat (Claude, ChatGPT and other chat apps), Use with agents (Claude Code, Codex, Cursor and other AI agents, open by default) and Use in your own app (the MCP server itself, for your own code).

Signing in

• Under Use with agents, copy the setup for your agent and paste it into Claude Code, Codex, Cursor, OpenCode or any other AI agent. It holds only the server’s name and address, and points the agent to AI agent setup, where it finds the sign-in steps for its own client and adds the server itself.
• When the agent connects, or when you start its sign-in (in Claude Code: /mcp, then the server, then Authenticate), a Nevazu page opens in your web browser. Sign in to Nevazu first if it asks.
• The page names the app and where it will return. Choose whether the agent may act or only look, and how long the sign-in lasts: 1, 7, 30 or 90 days, 7 by default. Then choose Allow. If connections are off for the browser, allowing turns them on.
• The agent stays connected until the time you chose and renews its access by itself. After that, sign it in again the same way.

Nevazu emails you each time an agent signs in. If the page says the app’s name is not verified and you did not just start a sign-in from your own agent, choose Deny.

To configure an agent yourself, open Set it up by hand: it shows the command or configuration file for each client, already filled in for that browser.

Using it in a chat app

Chat apps that can add a remote MCP server, such as Claude (claude.ai and Claude Desktop) and ChatGPT, connect the same way. Under Use in chat, copy the server URL and follow the steps for your app: in Claude, Customize → Connectors → + Add → Add custom connector; in ChatGPT, Plugins → + → Add custom MCP server, with OAuth, then Create as a plugin. The app opens the same Nevazu approval page. Claude: custom connectors ↗ ChatGPT: add an MCP server ↗

Using it from your own app

Use in your own app gives the server URL and how it authenticates, for a script, an app or a self-hosted AI agent of your own. The server speaks MCP over Streamable HTTP and uses OAuth 2.1 with PKCE (MCP authorization): a request without a token gets a 401 whose WWW-Authenticate header names the protected resource metadata, which in turn names Nevazu as the authorization server. Clients register with Dynamic Client Registration or a Client ID Metadata Document, ask for browser:control or browser:view, and send the access token as Authorization: Bearer.

Access keys

Access keys are no longer offered: new connections sign in. Keys created before keep working until they expire or you revoke them, and are listed with the browser’s connections.

Disconnecting an agent

The browser’s Settings → Agent access lists its connections, with who is driving, when each ends and when it was last used. Disconnect one there at any time, or open Agents in the dashboard’s sidebar: it shows every agent of all your browsers, which one is working now and who has control, and lets you pause an agent, hand control back, or disconnect agents one at a time or all at once. A disconnected agent loses access at once, even mid-task. Deleting the browser or signing out on all devices also disconnects its agents.

A key or a sign-in works only for the browser it was made for. Treat a key like a password: anyone who has it can see and act in that browser with your sign-ins.

Who is driving

While an agent is connected, the live view says who is driving: a green outline around the picture while the agent is in control, a blue one while you are, and the robot button in the toolbar, whose menu holds the choices below. A short notice appears when control changes; it closes by itself. When the agent asks for control while you have it (it tried to act, or it is waiting for you), a speech bubble points at the robot button: tap or click the button to hand control back, or choose Not now to keep working; the bubble also goes away once the agent stops asking. Any click, key, scroll, paste or dropped file of yours takes control from the agent before it is carried out: the agent’s remaining steps are cancelled, keys it was holding are released, and it is told it lost control. Moving your mouse without clicking does not take over, so the page the agent is looking at does not change under it.

Once you have taken over, the agent stays paused until you choose Hand control back, in the robot button’s menu, on the Agents page or in Settings. Closing the live view, reconnecting, or the agent reconnecting does not hand control back. When you hand it back, the agent takes a fresh screenshot before it continues; nothing it had queued before is resumed.

Privacy pause goes one step further: besides pausing input, no screenshot reaches the agent until you end the pause. Screenshots the agent already received cannot be taken back.

Screens and coordinates

The agent sees the whole remote display, including tabs, the address bar, menus and file dialogs, scaled to a size models read reliably; the coordinates it sends are mapped back to the real screen. Every action names the screenshot it was decided on: if your window changed the remote screen size or you took control since, the action is refused and the agent takes a new screenshot. Resizing never takes control away from it.

Files

To give the browser a file from your computer, the agent asks for a one-time upload address and sends the file with curl, which comes with macOS, Linux and Windows; small files can also be sent directly. Uploading only copies a file to the browser’s temporary storage; attaching it to a page (into Chrome’s file dialog, by paste or by drop) needs the agent to have control at that moment. Temporary files are removed when the agent’s session ends.

What the agent can do

The agent has the same reach as you at the keyboard and mouse: every website you are signed in to, including anything those accounts allow (reading, sending, buying, deleting). Give an agent only tasks you would do yourself, watch it in the live view, and prefer a separate browser with its own sign-ins for agent work. Every request the agent makes (a screenshot, an action, a file, a status check) counts as activity, so a browser an agent is working in does not go to sleep, even when nobody is watching. An idle connection does not keep it awake: once the agent stops making requests, the browser sleeps on its usual timer.