AI control

AI control lets an agent use OpenArch itself: open pages, click, fill in forms, take screenshots, and read a page’s console and network requests, in your real tabs and signed in as you. It is meant for Claude Code, and works for anything else that speaks Chrome’s DevTools protocol, such as Playwright or Puppeteer. It is off by default, because while it is on, any program on the Mac can use it.

Turn it on

1. Open Settings › KI and find the section KI-Steuerung (AI control).
2. Turn on KI-Steuerung erlauben (Allow AI control).
3. Leave Port at 9222, or enter another.
4. OpenArch opens the port at launch, so the settings now say “Wird nach einem Neustart von OpenArch aktiv.” (Takes effect after a restart). Click Jetzt neu starten (Restart now).

Afterwards the settings show “Aktiv auf 127.0.0.1:9222”. Turning it off, or changing the port, also takes a restart.

The antenna

While the port is open, an orange antenna sits at the top of the sidebar, next to the button that hides it. Resting the pointer on it shows “KI-Steuerung aktiv: Chrome DevTools Protocol auf 127.0.0.1:9222”. When the sidebar is too narrow for it, an orange dot on the sidebar button takes its place.

Connect Claude Code

Settings › KI › Mit Claude Code verbinden (Connect with Claude Code) shows the command to run once in Terminal, with a button Kopieren:

claude mcp add openarch -- npx -y chrome-devtools-mcp@latest --browser-url=http://127.0.0.1:9222

It adds Google’s MCP server chrome-devtools-mcp to Claude Code under the name openarch, pointed at OpenArch’s port. The port in the command follows the Port field. The server is started with npx, so it needs Node.js on the Mac.

Note

claude mcp add adds a server for the folder you run it in. For Claude Code to have it in every folder, add --scope user: claude mcp add --scope user openarch -- npx -y chrome-devtools-mcp@latest --browser-url=http://127.0.0.1:9222

Agent chats and page chats in OpenArch do not need this command: while AI control is on, OpenArch gives every run of theirs the same server itself.

What an agent can do with it

Through the port, an agent works in the browser the way you do: it opens pages and tabs, clicks, types and fills in forms, takes screenshots, and reads the console and the network requests. It works in your real tabs, in your spaces’ profiles, signed in wherever you are signed in. This is what makes it useful for testing a page you are building locally, or for a task on a site that needs your login.

In OpenArch’s own chats, these steps show as “Browser: …”. Unless a chat is set to Alles erlauben (Allow everything), its first browser step is refused and shows “Browser-Steuerung nicht erlaubt” with Erlauben und weiter (Allow and continue), which allows OpenArch’s browser control for that chat. A page chat is told to use the browser only when you ask it to do something in a tab.

Other tools that speak the DevTools protocol, such as Playwright or Puppeteer, connect to the same address, http://127.0.0.1:9222.

Keep it off when you do not need it

Caution

While the port is open, any program on this Mac can control the browser through it, not only Claude Code, and without asking: including tabs where you are signed in. It listens on 127.0.0.1 only, so it cannot be reached from other devices. Turn AI control on when you need it and off again afterwards.

The antenna in the sidebar is there so that you see when the port is open.

What Claude gets and what it may do

With AI control, an agent can see whatever your tabs show and do what you can do in them. Agent chats and page chats get OpenArch’s browser server only while AI control is on, and what a chat may do without asking is set per chat.

OpenArch does not give the server to the Morning Brief or to tidying tabs. If you added it to Claude Code with --scope user, though, the Morning Brief sees it like any other connector and may use the tools whose names say they read, such as listing the tabs.

How the runs work in detail: How OpenArch talks to Claude.