Why the Cloud Browser Fails When Your Own Browser Is Fine
When cloud browser not working chatgpt is the problem in front of you, the cause sits in one of four layers: plan and region availability, workspace permissions, per-site blocking, or a stuck remote session. The browser runs on OpenAI’s machine, so clearing your own cache changes nothing. Work the layers in order — plan, admin policy, site permission, session handover — and most failures clear in minutes.
📌 TL;DR Executive Summary
- Core Takeaway: Cloud browser runs on a separate machine in OpenAI’s cloud with its own cookies, sign-ins and per-site permissions. Your local cache, extensions and browser settings never travel with it.
- Key Risk/Challenge: Most reports are a website blocking signed agent traffic or a policy gap — and a saved Deny from one chat keeps blocking that site until you start a new chat.
- Recommended Solution: Check the status page, open a fresh chat, review Settings > Cloud browser, take over the session when it pauses, and keep your own work logins in separate browser profiles so sessions never mix.
What the Cloud Browser Actually Is, and What It Ignores
In ChatGPT Work, cloud browser gives ChatGPT its own browser on a separate computer in the cloud. It loads pages, clicks buttons and fills forms there, then reports back into the conversation. That machine keeps its own cookies, its own browser data and its own signed-in sessions.
The consequence is practical: it does not use your tabs, history, saved passwords or extensions. When cloud browser not working chatgpt is what you are chasing, the three fixes people try first — clearing cookies for chatgpt.com, disabling extensions one at a time, switching your VPN — have no path to the remote browser. They repair the ChatGPT website in your tab, not the agent’s machine. If your chat interface loads fine and only the task fails, you are debugging the wrong computer.
Availability is the first gate. Cloud browser ships on paid ChatGPT plans in supported regions and is not part of Free or Go. OpenAI also notes that availability varies during rollout and by workspace permissions, which is why two colleagues on the same plan can see different menus. The official cloud browser documentation is worth reading before you assume a bug.
The four layers, and how each one looks when it breaks
| Layer | What it controls | How the failure looks |
|---|---|---|
| Plan and region | Whether cloud browser exists for your account at all | The option is missing entirely; no error, no trace |
| Workspace policy | Admin switches for browser use, site access, file transfers, history access and saved approvals | The settings page loads but controls are greyed out or absent |
| Site permission | Always ask, Auto approve, Always allow, plus per-site allow and block rules | The task stops before loading, or a site you allowed earlier still refuses |
| Remote session | The cloud machine’s browser state, sign-ins and runtime | Page half-loads, the agent asks you to take over, or the session crashes |
Why a website can tell it apart from your visit
Cloud browser signs its outbound HTTP requests with HTTP Message Signatures, the mechanism defined in RFC 9421. Each request carries Signature, Signature-Input and a Signature-Agent header set to https://chatgpt.com, with the matching public keys published at a well-known directory on chatgpt.com. CDNs verify that signature instead of guessing from behaviour, which is what the cloud browser allowlisting process is built around.
The recognised entries already exist. Cloudflare treats the traffic as bot tag chatgpt-agent with detection ID 129220581, Akamai lists it as ChatGPT Agent under its AI bots category, HUMAN does the same, and Vercel allows it with no configuration. So when a site refuses the agent, it is usually a deliberate rule rather than a misread.
If you want the general mechanics of remote browsing before you go further, this explainer on what a cloud browser server is covers the architecture in plain terms.
Quick Wins, Cheapest First
Before you change any setting, spend two minutes eliminating the obvious. When cloud browser not working chatgpt turns out to be a stale approval or a paused handover, these six steps catch it. They are ordered by effort, not by likelihood.
- Check status.openai.com. ChatGPT, the API and Codex are reported separately, and you can subscribe to incident emails. A Downdetector spike often appears before the official post, so treat it as a second opinion rather than a verdict.
- Start a new chat. A saved Deny in one conversation can keep blocking a site even after you have allowed it generally. A fresh chat drops that sticky decision, and it is the fastest fix for “it worked yesterday.”
- Re-check the permission level. Settings > Cloud browser offers Always ask, Auto approve and Always allow, plus per-site rules. If someone switched you to Always ask, every navigation now waits for a click you may not have noticed. OpenAI describes Always allow as not recommended, so use it narrowly.
- Try an unrelated site. Loading a different page separates “this site blocks the agent” from “the session is broken.” One works, one fails, and you now know which half of the problem you own.
- Take the handover. When the agent pauses and offers you control, accept it, or request a link to drive the browser directly. Many apparent freezes are a task waiting on a human.
- Confirm eligibility. Paid plan, supported region, and a workspace that has not restricted browser use. If any of those is missing, retrying will not help.
/home/agent/.codex/sessions tells support you hit the remote runtime; “it just hangs” tells them nothing they can search.
When the Page Never Loads Because the Site Said No
Site-level blocking is set by the website, not by ChatGPT. If cloud browser not working chatgpt happens on one domain while other pages load normally, you have found the layer. Each operator decides whether agent traffic is welcome, and OpenAI’s guidance to users is direct: try a different website, or open that page yourself in your own browser. There is no user-side toggle that makes a site accept the agent.
Here is what is actually happening underneath. The site’s edge, or its bot-management vendor, holds a rule that either allows or denies the chatgpt-agent signature. If the rule is missing, the request falls through to a default deny. If the rule exists but an intermediate proxy strips Signature, Signature-Input or Signature-Agent, verification fails and the request arrives looking unsigned. OpenAI’s allowlisting notes tell operators to check both header preservation and the exact bot name or detection ID.
If you run the site, that is your fix list. If you do not, accept the boundary and move the task into your own browser. Retrying through a proxy or a different network does not change another company’s rule, and trying to slip past it is exactly the kind of workaround that ends with a suspended account.
When the Task Stalls, Asks for a Sign-In, or Hands You Control
A pause is often a feature. Cloud browser stops when it needs your input, your sign-in or your confirmation, and OpenAI advises checking results and sources before you rely on them. If the conversation looks frozen at a login screen or a confirm button, look for the take-over prompt rather than refreshing the tab.
Sign-in support is the murkiest area, and the documentation disagrees with itself. The cloud browser help article describes a secure sign-in form, while the allowlisting page, citing HUMAN’s documentation, states that at launch the agent cannot sign in to websites or complete payments, and that permission settings do not change this. Treat any task that depends on an authenticated session or a checkout as unreliable inside cloud browser. Do the login step yourself, then hand back a page that does not need credentials. If you are using the agent around client accounts, how to use the cloud browser in ChatGPT safely covers the permission boundaries in more detail.
Verification steps belong to the same category. Complete them as yourself, in your own browser, through the normal flow. A verification prompt is the platform asking for you, and the correct response is to be there.
When the Cloud Browser Setting Isn’t There at All
A missing menu is not automatically a bug. Three things remove it: an account that is not on a paid plan, a region where the rollout has not reached you, and a workspace administrator who has switched off browser use or site access. Inside a workspace, an admin can also restrict file transfers, history access or saved approvals, which greys out individual controls rather than the whole page.
Work through it in writing. Ask your admin whether Browser Use is enabled for your seat and whether site access has its own policy. If you are the admin, check the workspace settings before you file a support ticket. If you are on a personal paid plan in a supported region and the option is still absent, that is worth reporting — but send the plan name, region and exact timestamp rather than “it doesn’t work.”
Reading the Error: node_repl, bwrap and the Atlas Confusion
Older tutorials will send you down the wrong path. OpenAI discontinued the standalone Atlas browser on July 9, 2026 and folded its functions into ChatGPT Work, a task agent backed by Codex, while the older chat-only desktop app became ChatGPT Classic. If a guide tells you to fix Atlas settings, it is describing a product that no longer exists.
Recent crash reports carry their own vocabulary. A community thread from late October 2026 describes cloud browser failing to initialise with a node_repl kernel exit and a bwrap read-only filesystem error at /home/agent/.codex/sessions. The same user reproduced it across browsers, networks and physical machines while ordinary code and shell commands kept working in the same conversation. That pattern points at the remote runtime, not at anything you can change locally.
One report is not an outage, so treat it as a signature to recognise rather than a diagnosis. If you see a path inside the agent’s own filesystem, stop reinstalling your browser and file the report with the error text, the timestamp and your plan.
Cloud Browser vs Desktop App Browser vs the Extension
Half of the “cloud browser broken” tickets are really a mismatch between three different tools. They run in different places and hold different state.
| Cloud browser | Desktop app built-in browser | ChatGPT browser extension | |
|---|---|---|---|
| Where it runs | A separate machine in OpenAI’s cloud | Inside the ChatGPT desktop app on macOS or Windows | Your existing Chrome profile, on the web |
| Whose sessions it uses | Its own cookies and sign-ins | Its own browser state, not your Chrome profile | Your tabs, sessions and extensions |
| Install required | None | The desktop app; open it with Cmd/Ctrl+Shift+B | The Chrome Web Store extension |
| Best for | Delegated reading and clicking that needs no account | Work inside the app with autofill, downloads and extensions | Tasks that genuinely depend on being signed in as you |
If your task needs your logins, cloud browser is the wrong tool and the desktop browser or extension is the right one. If the task has to run while your laptop is closed, the reverse is true — and the cloud browser isolation guide for remote work explains how those remote sessions should be scoped once more than one person is involved.
How to Keep It From Recurring
Most repeat failures come from configuration drift, not from OpenAI. If cloud browser not working chatgpt keeps landing on your desk every few weeks, the fix is a short, deliberate routine rather than another round of reinstalls.
- Keep one task per chat. Saved approvals and denials attach to conversations. A fresh chat per job stops yesterday’s Deny from sabotaging today’s run.
- Audit your allowlist monthly. Per-site rules accumulate. Remove destinations you no longer use and keep Always allow for the two or three you truly need.
- Do logins and checkouts yourself. Assume cloud browser will not carry an authenticated session, and design tasks that finish without one.
- Subscribe to status updates. An email subscription to incident posts beats refreshing the tab on a Monday morning.
- Separate your own working sessions. If you manage several client, brand or storefront accounts, the recurring cause is mixed sessions on your side — the wrong cookie jar answering the wrong login prompt.
Run Cloud Browser Not Working Chatgpt in the Cloud With Send.win
Send.win’s cloud browser runs isolated browser profiles on a server in the EU — open a clean session from any device without installing anything:
- Instant cloud sessions – launch an isolated browser in seconds, no local install
- Isolated profiles – separate cookies, storage and fingerprint settings per session
- Cloud sync & profile sharing – pick up the same profiles on the desktop app (Windows, macOS, Linux) or share them with your team
- Residential proxies included – on Pro and Team, with timezone and locale matched to the proxy
You can try it right now: the Send.win demo browser opens an isolated cloud session directly in this browser tab. Paid plans start at $19/month, with a 7-day money-back guarantee on your first card purchase — see pricing.
That last point is where a profile manager earns its place. Sendwin Browser keeps each working identity in its own profile: separate cookies, storage and signed-in sessions, its own fingerprint settings, and its own proxy so timezone, locale and geolocation follow the exit IP. Switching between client A and client B means switching profiles, not logging out of one account and into another in the same tab. For teams, a profile shared with a paid teammate opens already signed in, so no password changes hands. Use it only with accounts you own or are authorised to manage, and follow each platform’s terms.
If your mixing problem shows up inside a single browser instead, the same logic applies at tab level — this guide to Firefox containers not working walks through container isolation when it refuses to behave.
🏆 Send.win Verdict
ChatGPT’s cloud browser is a capable delegated reader for public pages and a poor fit for anything behind a login — which is exactly where your own accounts live. Send.win sits on the other side of that line. It gives you isolated browser profiles with their own cookies, storage, sessions, fingerprint settings and proxy, in a native desktop app for Windows, macOS and Linux, or as cloud profiles running in Helsinki from any device. When the real reason a task fails is that three client logins are fighting over one cookie jar in your own browser, separate profiles fix the cause instead of the symptom.
See Send.win plans — Pro at $19/mo (or $83.88/yr, about $6.99/mo) with 150 profiles and 6 seats, Team at $49/mo with 500 profiles and 16 seats, plus a free cloud demo if you want three minutes with no signup.
Frequently Asked Questions
Why is ChatGPT’s cloud browser not loading?
Check three things in order: whether status.openai.com lists an incident, whether your plan and region include cloud browser, and whether the target site blocks agent traffic. If the conversation stalls rather than erroring, the agent may simply be waiting for your input or a handover.
Does cloud browser work on the free ChatGPT plan?
No. It is available on paid ChatGPT plans in supported regions, and Free and Go are excluded. Availability can also vary during rollout and by workspace permissions, so a colleague on the same tier may see it before you do.
How do I allow a site in the cloud browser settings?
Open Settings > Cloud browser, choose a permission level — Always ask, Auto approve or Always allow — then add per-site allow or block rules. If a previously denied site still fails, start a new chat: the saved decision belongs to the old conversation.
Why does the cloud browser stop at a sign-in page?
The agent runs on a separate machine with its own session, so it does not inherit your logins. OpenAI’s help centre describes a secure sign-in form, while its allowlisting page says the agent cannot sign in to sites or complete payments at launch. Treat login-dependent tasks as unreliable and handle that step yourself.
Can I take over the cloud browser manually?
Yes. When the agent gets stuck it asks you to take over, and you can request a control link directly as well. Taking control is faster than re-prompting and it keeps the session state intact.
Is cloud browser the same as the desktop app’s built-in browser?
No. The desktop app’s built-in browser runs locally inside the app with its own browser state and supports autofill, passwords, extensions and downloads. Cloud browser runs on OpenAI’s machine and holds its own cookies and sessions.
Why does the cloud browser forget my logins?
Its cookies and sessions live on the remote machine and are not your browser’s cookies, and sessions can reset between runs. Any workflow that depends on staying signed in is better handled in a local browser profile you control.
Will clearing my cache or disabling extensions fix it?
Almost never, and that is the most common wasted hour here. Those actions change your local browser, while the failure usually happens on OpenAI’s machine or at the website’s edge. Start with the status page, a new chat and the permission settings instead.