Let Agents Drive the Browser

Your AI agents can operate the Browser panel you already use — same tabs, same history, same signed-in sessions.

Headless browser automation starts from nothing: no cookies, no session, no extensions, and a login wall on the second page. So the agent asks you for credentials, or gives up and describes what it would have done.

1DevTool hands agents the browser you already have open instead. Same tabs, same navigation history, same signed-in sessions — the agent works where you work.

An agent operating the Browser panel

What an agent can do

Any agent connected through 1DevTool MCP can:

  • Navigate — open a tab, select a tab, go to a URL, go back or forward, reload, close a tab
  • Read — take a structural snapshot of the page, read console output, capture a screenshot of the current tab
  • Act — click a control, type text, choose an option from a select, press a key, wait for the page to change

That covers most of what "check whether the fix works" actually requires: load the page, click through the flow, read the console, screenshot the result.

What it can't do

Automation stays scoped to an explicit project and tab. An agent doesn't roam your whole browser because you asked it about one page.

Password and file inputs are protected — an agent can't type into a password field or hand your file picker a path.

Point an agent at it

Ask for the browser directly and the agent uses this panel:

Use the 1DevTool browser MCP: open localhost:3000/pricing, click the annual toggle, and tell me what the console logs.

Agents that would otherwise reach for their own Chrome session — Codex in particular — route to the visible Browser panel instead of launching an external window or spawning another agent terminal.

An agent routed to the in-app Browser panel

Page references survive navigation

A snapshot gives the agent element references. When the page navigates, those references are refreshed rather than silently pointing at a page that no longer exists — so a multi-step flow doesn't fall apart on the redirect after login.

Big pages stay responsive

Screenshots and large page payloads go to the agent in full, while the desktop interface processes only a compact preview. The agent gets the complete result; your window doesn't freeze rendering it.

Watching what it did

Every call shows up in MCP Control → Activity with its status, duration, and a safe preview of the result — see MCP Control. If browser automation is more access than you want in a given session, turn the Browser group off there and the agent is told the tool is unavailable.

Next