Codex Out of Quota? Connect ChatGPT Web to Codex (2026 Guide)
📚 Table of Contents (click to jump)
- 1. What Is It? (30-Second Overview)
- 2. Install Codex Web GPT
- 3. Log in to ChatGPT and Check the Connection
- 4. Load Models into Codex
- 5. What It Can and Cannot Do
- 6. Troubleshooting First
- 7. How to Uninstall Cleanly
- FAQ
1. What Is It? (30-Second Overview)
It takes about 10 minutes to install if things go smoothly. You need only two things up front: Codex installed (desktop app or CLI) and a ChatGPT account that can sign in to the web app. No API Key, no Chrome, and no configuring Node or Bun yourself.
The project is ChatGPT Web for Codex (the desktop program this guide installs is called Codex Web GPT), MIT-licensed, around 6.8k stars.
In one sentence: it makes ChatGPT web models appear in the Codex model picker.
Before: Codex → Codex built-in models → consumes Codex quota
After: Codex → ChatGPT web models → uses ChatGPT quota (no Codex quota spent)Let's nail one thing down first, because it prevents most later confusion:
It does not convert ChatGPT quota into Codex quota. Both sides still count separately — you simply gain an extra usage path.
So the right expectation is "one more option," not "double the quota." When your Codex quota runs dry, there's still a way to keep working.
This is a third-party project, not an official OpenAI feature. Under the hood it's browser automation — it runs a local forwarding service and hands your requests to a temporary ChatGPT conversation bound to the current task — so a web redesign can break it. That's the premise you should accept before deciding to use it.
Before you try it with a main account, three risks are worth knowing:
- A web redesign will break it. When the interface structure no longer matches it reports an explicit error rather than pretending to keep running.
- The account side may see temporary rate limiting. The project caps simultaneously open tabs at 5, but the official line is that this "does not represent a safe concurrency number," and some accounts get limited with only two tasks open. Start conservatively.
- The login state the launcher saves is a sensitive credential. Don't share the config file with others, and don't use it on an untrusted computer.
To play it safe, run it on a secondary account for a while first.
2. Install Codex Web GPT
Open PowerShell and run:
irm https://github.com/miuuyy/codex-chatgpt-web/releases/latest/download/install-launcher.ps1 | iexOn macOS / Linux, open a terminal and run:
curl -fsSL https://github.com/miuuyy/codex-chatgpt-web/releases/latest/download/install-launcher.sh | shDon't feel like memorizing commands? Hand the repo link to your Codex and let it install for you:
https://github.com/miuuyy/codex-chatgpt-web install this project for me
and tell me how to use it. Go ahead.Once it's installed, open Codex Web GPT from your launcher list.
⚠️ The project's builds aren't currently code-signed, so macOS Gatekeeper or Windows SmartScreen may throw a warning once — allow it to open. The install script verifies the official published SHA-256 manifest before installing.
3. Log in to ChatGPT and Check the Connection
Open Codex Web GPT and you'll see three checks.
Step 1: log in to ChatGPT.
This step is not asking you to enter an API Key — it's signing in directly to the ChatGPT web account you normally use.
Here's the catch: the login must happen in the launcher's own built-in browser window. If you open a separate Chrome and sign in there, the login state won't carry over. Don't close or switch away from the window during login either — wait until the temporary conversation's input box appears before continuing.
Step 2: run the Browser Smoke Test.
This is a browser smoke test: it uses one simple exchange to confirm the launcher can drive ChatGPT's web interface and retrieve an answer.
Only when it passes is step 1 truly complete. If it fails, don't rush to the next step — check the error table in section 6 first, or the Install Models step later is guaranteed to install nothing useful.
When all three checks turn green, this section is done.
4. Load Models into Codex
Step 3: click Install Models.
What this does: it writes a route into the Codex configuration so the model list gains ChatGPT Web tiers.
🔴 The step most likely to fail and most likely to be overlooked: after installing you must fully quit Codex and reopen it. Closing the window, starting a new task, or signing out all don't count — if the process doesn't end, it keeps using the old model list and the models simply never appear. Codex CLI has to be quit too.
Also, the launcher must stay open — model requests are forwarded through it.
Restart Codex, and the model picker should show the new tiers:
ChatGPT Web — instant
ChatGPT Web — medium
ChatGPT Web — high
Which tiers you can install depends on your account's entitlements:
| Your account | Options that may appear |
|---|---|
| Free / Go | Usually only basic tiers like Luna and Think |
| Plus | instant / medium / high |
| Pro | Plus the Extra High and Pro tiers on top of the Plus set |
💡 After selecting a ChatGPT Web model, the Effort / Speed options still appear in the Codex interface, but changing them won't actually switch you to a different ChatGPT model — to change tier, change the entry in the model picker.
Seeing the model list means the install succeeded. From there it works just like using Codex normally.
Two boundaries worth knowing right after install:
- ChatGPT Web missing from the model list? Don't reinstall repeatedly — go to section 6 and troubleshoot.
- It cannot read files on your computer and cannot execute commands (Codex shows a warning, which is normal). File access requires configuring the advanced full MCP mode, which is a separate article's worth of material.
5. What It Can and Cannot Do
Knowing the boundaries matters more than knowing the features. Its main battlefield is "thinking work, not hands-on work": reading code, analyzing errors, code review, discussing implementation approaches — none of these depend on local file operations, so switching over is the best value.
Two things you shouldn't expect from it, on the other hand:
- File reading/writing and command execution. With only the launcher installed it just holds a conversation and can't touch anything on your computer (file access requires the advanced full MCP mode, outside this article's scope).
- Codex's native slash commands. Things like
/modeland/compactin the CLI are the outer Codex's own features, not capabilities of the ChatGPT web app. Which ones still work after switching to ChatGPT Web depends on the outer Codex — don't assume. (The project's own subcommands, such ascodex-chatgpt-web subagents status, are outside this scope.)
As for when to switch, the criterion is simple:
- Codex quota healthy → use Codex's built-in models as usual.
- Codex quota tight → switch to ChatGPT Web and hold out for a while.
6. Troubleshooting First
The general troubleshooting principle: don't reinstall repeatedly. One accurate error message plus one log is worth more than ten reinstalls. Settings → Run doctor in the launcher runs an end-to-end self-check, and Activity → Export safe log exports a redacted log.
| What you see | Most likely cause | What to do |
|---|---|---|
| No ChatGPT Web in the model picker | Codex wasn't truly restarted | Fully quit Codex (including background processes and CLI) and reopen it, keeping the launcher running; if that fails, click Repair Codex setup once, then run Run doctor |
| The three checks stall at step 3 but the browser test is green | Step 3 already succeeded | Clicking step 3 again won't fix browser or model request problems — go straight to the specific error |
| Can't get past the login page | Login must happen in the launcher's own window | Complete the launcher's login flow and wait for the temporary conversation input box; if there's a "try another way to sign in," use it first rather than repeatedly deleting the browser profile |
| Smoke test fails | Usually ChatGPT's page structure changed, or the account is signed out / rate limited | Update to the latest version → confirm you can open a normal temporary conversation with no rate-limit prompt → run it again; if the same error repeats, stop retrying and export the log for an issue |
openai_base_url changed after setup | Another tool is also managing Codex routing | Check whether routing tools like CC Switch, OmniRoute, or Codex++ are installed. Only one program can manage openai_base_url at a time: to use this project, turn off the other tool's proxy mode, click Repair Codex setup, then launch Codex directly; to switch back, first use Settings → Remove Codex integration |
| Repeated Reconnecting / stream disconnected | This is only the result boundary, not the cause | Read the last detailed error; retry once in a new task to see if it reproduces; run Run doctor and export the log |
| ChatGPT says the account is temporarily limited | Account-side rate limiting | Stop retrying immediately and wait for cooldown. You can set max_concurrent_threads_per_session to 1 under [agents] in ~/.codex/config.toml (if the table already exists, change only that item — don't create a new one) |
| Image generation runs halfway with no image | Known limitation | Image generation inside ChatGPT web conversations isn't a currently supported turn type |
Errors related to full MCP mode aren't in this table — that's advanced configuration for a separate article.
🔐 Before filing an issue, don't paste raw logs, cookies, browser storage, API Keys, Tunnel IDs, full prompts, or tool output containing private data. Use Activity → Export safe log.
7. How to Uninstall Cleanly
Order matters here, and step 1 must come first:
- Open Settings → Remove Codex integration and let it restore Codex's routing to its pre-install state;
- Fully restart Codex;
- Quit the launcher and uninstall the program normally through your system.
Why the order matters: if you uninstall the program first, Codex keeps pointing at a local route that no longer exists, and you'll hit a pile of baffling connection errors.
FAQ
Q: Does this turn ChatGPT web quota into Codex quota? No. It only adds a ChatGPT Web usage path inside Codex, and the two quotas stay independent. What it gives you is the choice not to stare only at your Codex quota.
Q: It's installed but ChatGPT Web isn't in the model list. Nine times out of ten Codex wasn't truly restarted. Fully quit it (including background processes and CLI) and reopen it, keeping the launcher running.
Q: Does it need an API Key? No. It uses the ChatGPT web account signed in inside the launcher. Only the advanced full MCP mode requires creating a Tunnel and API Key in your OpenAI account — creation is free and doesn't consume model quota.
Q: Can it use Codex's file and terminal tools? Not by default; with only the launcher installed it just holds a conversation. File access requires configuring the advanced full MCP mode separately, and accepting the write-permission risk it introduces — that's another article.
Q: Are there account risks? The project states it doesn't bypass authentication or hand out new accounts. But it's unofficial usage: a web redesign breaks it, and the account side may see temporary rate limiting. With a main account, read up first, or try a secondary account.
Q: macOS says the developer cannot be verified? Normal. The project's builds aren't code-signed yet, so Gatekeeper / SmartScreen blocks it once — allow it to open. The install script verifies the official SHA-256 manifest.
Looking back at that opening line — what this project solves is the mismatch of "Codex quota almost gone while the ChatGPT web subscription sleeps." At least now you don't have to ration Codex quota while leaving ChatGPT quota unused.