Debug console and network errors with an AI agent

To debug a page with an AI agent, connect the agent to your Chrome with Agent Chrome and ask it to reproduce the bug and read the console and network log. It calls read_console_messages for errors and read_network_requests for failed calls, then reports which request or script broke.

The agent reads the same tab you would open DevTools on, with your login, so bugs that only show up for your account can be reproduced.

The problem: you are the copy-paste layer

The usual loop is: reproduce the bug, open DevTools, copy the red error, paste it into the chat, then copy the failing request, paste that, and repeat when the agent asks for more. Agent browser tools that start a clean browser cannot reproduce bugs that need your session, your feature flags or your data.

With Agent Chrome the agent does the reproducing and the reading itself.

Walkthrough: find why a save button does nothing

Ask your agent:

Use agentchrome to open http://localhost:3000/settings, click "Save changes", and tell me why nothing happens. Check console errors and failed network requests.

The agent will call:

1. tabs_context to get its tab group.
2. navigate with url: "http://localhost:3000/settings". The console and network capture starts from the first time the agent touches the tab.
3. find with query: "Save changes button" returns a ref and coordinates.
4. computer with action: "left_click" and the ref.
5. wait_for with textGone: "Saving..." and timeoutMs: 5000 to let the request finish, or fail after the timeout.
6. read_console_messages with onlyErrors: true returns uncaught errors and console.error lines. Use pattern to filter by regex, limit to cap the count, and clear: true to reset the log between runs.
7. read_network_requests with urlPattern: "/api/" and includeBodies: true returns each request's URL, method, status, type and timing. With includeBodies, XHR and fetch responses include up to 10 KB of body, which is where an API error message usually sits.

A typical report from the agent: "The click sent PATCH /api/settings and got a 422. The response body says email: must be unique. The console shows Uncaught TypeError: Cannot read properties of undefined from the error handler, so the UI never shows the message."

Go further with javascript_tool

When logs are not enough, the agent can inspect the page itself. javascript_tool runs JavaScript in the page's main world and returns the value of the last expression. It supports await, works on pages with a strict Content Security Policy, and reports a thrown error as an error result.

Examples you can ask for:

Check what localStorage.getItem("token") holds and whether the token has expired.

Call fetch('/api/me') from the page and show me the JSON.

The tool call is javascript_tool with text: "await fetch('/api/me').then(r => r.json())". Ask the agent to return small values. Huge objects are not a good return value.

Check what the user would see

Console and network are half the picture. Add visual and state checks:

• computer with action: "screenshot" shows the error toast, or the lack of one.
• screenshot with fullPage: true captures the whole page.
• read_page returns the accessibility tree, so the agent can check whether an error message exists in the DOM but is hidden.
• get_page_text reads the main content.

If the bug depends on timing or on a dialog, note that alert, confirm and prompt are reported to the agent and handle_dialog answers them, and HTTP errors such as 404 and DNS failures are reported by navigate.

Record the bug

Once the agent can reproduce it, ask for proof you can attach to the issue:

Reproduce the bug again with a recording on, then stop and give me the link.

The agent uses record with action: "start" and action: "stop". See Record a browser video or demo with AI.

Tips and limits

• The log starts when the agent first touches a tab. Errors that fired earlier, before the agent attached, are not there. Reload with navigate and url: "reload" to capture a page load from the start.
• Filter early. pattern, onlyErrors and urlPattern keep the result small, so the agent spends less context on noise.
• Response bodies are capped. includeBodies returns up to 10 KB per XHR or fetch response.
• This is not a performance profiler. For traces and Lighthouse-style audits, Chrome DevTools MCP is the better tool. See Agent Chrome vs Chrome DevTools MCP.
• Chrome shows a "started debugging this browser" bar while the extension is attached. Cancelling it stops every agent.
• Use a test account if the bug involves writes. The agent can do what you can.

For a full end-to-end test around the bug, see Test your web app with an AI agent. If the failing page needs your login, read Let an AI agent work on sites you are logged into. The parameters of each tool are listed in the tool reference.

Frequently asked questions

Can an AI agent read my Chrome console?

Yes. With Agent Chrome, the agent calls read_console_messages and gets log, warning and error lines plus uncaught exceptions from the tab. You can filter by regex with pattern or ask for only errors with onlyErrors: true.

Can it see failed network requests?

Yes. read_network_requests returns URL, method, status, type and timing for each request, and you can filter by urlPattern. Set includeBodies: true to include up to 10 KB of the response body for XHR and fetch calls.

Can the agent run JavaScript on the page?

Yes. javascript_tool runs code in the page and returns the last expression. It supports await and works under a strict CSP.

Why is the console empty?

Capture starts when the agent first touches the tab, so earlier messages are missing. Ask the agent to reload the page with navigate and url: "reload", then read the console again.

Does this replace Chrome DevTools MCP?

Not for performance work. Chrome DevTools MCP is strong at performance traces. Agent Chrome is the better fit when the bug needs your signed-in Chrome, or when you also want to click through the flow and record it.

Updated 2026-11-04 · Agent Chrome is open source (MIT) and operated by Circo, Inc.