Codex Opens Chrome but Can’t Read the Page: What to Check
A Chrome window opening is not evidence that Codex can read its page. Before reconnecting everything, identify which browser session you intended to use, then check the last step that actually worked: opening, connecting, reading, or saving.
I spent about ten days troubleshooting browser connections in a Windows Codex workflow. Chrome access, the built-in browser, and page reading and writing became available at different times. The problems appeared after an update, but I could not confirm it as the cause or one setting as a complete fix.
This is the check order I would use next time, for a desktop Codex workflow using external Chrome.
Start with what the result actually proves
| What you can verify | What it does not establish | Next useful check |
|---|---|---|
| You can see the page in Chrome. | Codex is using that Chrome profile or tab. | Identify the intended browser and profile. |
| The browser is available to select. | Access to this particular site is allowed. | Check the permission for the exact site. |
| Codex returns the current page title and a passage that matches. | It can edit the page or persist a change. | Only if needed, try a small, reversible draft change. |
| Codex reports that it entered text. | The website saved that text. | Reopen the intended draft and compare the saved content. |
“Connected” was too broad to tell me what to do next. A note such as “the current page was read; saving has not been tested” gives me a starting point without disturbing the parts that already work.
1. Choose the browser session before changing the connection
If the task depends on a site already signed in through your usual Chrome profile, keep that profile in view. Do not switch to an unrelated profile merely because it opens successfully: it may lead to a different account, document, or workspace.
The built-in browser is a separate session, not an automatic copy of your regular Chrome session. OpenAI documents this distinction in its Browser guide. A sign-in page in the built-in browser therefore does not, by itself, mean that your Chrome sign-in was lost.
My first question would be “Which window is this task supposed to use?” rather than “Should I sign out everywhere?” Clearing working sessions can turn a connection problem into an account-recovery problem without explaining the original failure.
2. Separate setup, browser selection, and site permission
The OpenAI extension guide separates setup from website access. Check these in order:
- In the ChatGPT desktop app, open
Settings > Computer Useand find Chrome. - Follow the official plugin and extension setup if it is incomplete. Confirm that Chrome shows
Manage. - In the Codex chat, select Chrome with an
@-mention and use the profile where the extension is installed. - Use
Manageto review access to the exact website. The browser toggle controls whether Chrome appears in the mention menu; it is not the site's permission switch.
Those are separate checkpoints. If Chrome is absent from the browser choices, investigate setup or visibility. If Chrome is present but one site is denied, keep the investigation at that site’s permission or policy. Do not treat an explicit denial as an invitation to find another route around it.
The aim is not to grant access to every site. It is to establish the narrow access required for the page you actually want Codex to use. An organization-managed restriction may require an administrator’s decision rather than a local connection reset.
3. Use a read-only check with an answer you can compare
Open a non-sensitive public page in the intended Chrome profile. Select Chrome for the task, then ask for a small, specific observation rather than a summary of the whole website.
Compare the answer with the page in front of you. A plausible explanation of the topic is not enough: the title and passage should match the current page, not something remembered from an earlier conversation.
If it does not match, do not proceed to a large edit. Record whether the failure occurred before browser selection, at site access, or while reading the page. That gives the next troubleshooting step a defined target.
4. Test saving only when the task needs it
Reading is enough for a research task. If the eventual job involves editing, use a disposable draft or another explicitly approved, reversible location. Ask for one short change, then inspect the saved draft yourself.
Keep the test out of a live post, payment form, message, or account setting. A test is useful only when its consequence is small and its result is unambiguous. Also check that a previous attempt did not already save before asking Codex to repeat it.
For me, the useful finish line is the saved result—not the agent’s statement that it clicked or typed. If the text is absent after reopening the intended draft, the write-and-save stage is still unverified even if reading worked.
5. Escalate the failed stage, not the whole workflow
If Chrome still cannot connect and the site is not blocked, use the extension guide's troubleshooting sequence. It covers app updates, browser and extension status, the active profile, and trying a new chat before later restart or reinstall steps. Check one thing at a time and keep other work safe before restarting anything.
A file-upload failure is a separate branch. The guide's Chrome setting Allow access to file URLs concerns uploading a file from your computer; it is not a general fix for an unreadable webpage.
Keep a short failure record
- The desktop app, browser, operating system, and versions you actually checked. Leave unknown versions unknown.
- The intended browser profile and task, without exposing personal identifiers.
- The last confirmed stage: open, connect, read, enter text, or save.
- The exact error with credentials, private URLs, and other sensitive details removed.
- The one change made and the result after that change.
I would stop at an explicit access restriction, an authentication step that needs my input, or an uncertain previous save. Otherwise, I would continue from the last verified checkpoint. That is a more useful outcome than another long explanation that leaves it unclear whether the page was ever read.
Before you reconnect again
- Is this the browser profile and account the task needs?
- Can Codex select the browser, and is this exact site allowed?
- Does a small read-only result match the current page?
- If editing is needed, did the intended draft actually retain the change?
What frustrated me most was spending time on explanations without getting back to the work. Next time, I want one clear line about the last verified step. The useful finish is not a window opening; it is the result I needed from that window.
Experience source: my Korean original, recorded October 6, 2026. Official references checked October 17, 2026: Browser extension and Browser. These steps are recommendations based on the linked documentation, not a new repair test. Current interface labels have not been independently verified.