Unofficial community examples for Codex web. Not affiliated with OpenAI. All trademarks belong to their owners.
Codex web is the browser-hosted surface of OpenAI's Codex coding agent, reached by signing in to ChatGPT and opening the Codex cloud page. It has no public API of its own that the ranking sources document, so this repository holds worked walkthroughs in markdown rather than scripts. Each one is a pattern I use when handing work to the agent from the browser, written so you can copy the prompt and adapt it.
If what you actually need is a finished page: Try Begin.sh - turn a prompt or a URL into a downloadable static site or Expo app.
| Section | What it shows |
|---|---|
| 1. A first cloud task | Scoping a single-change task so the result is easy to review |
| 2. Parallel tasks | Splitting independent work into several agent runs |
| 3. Long-running work | Writing a task that can run unattended |
| 4. Checking your plan and limits | Where limits come from and how to avoid surprises |
The examples/ directory is intentionally empty: there is nothing to run locally.
- A ChatGPT account. Sign in at chatgpt.com/codex/cloud.
- A plan that includes Codex. Per the Codex app announcement, Free and Go included Codex for a limited time and paid plans got doubled limits; confirm on the pricing page.
- The developer docs open in another tab for repository setup, which this repo does not duplicate.
No environment variables are involved; everything happens in the browser.
Start with something a reviewer can check in one sitting. A prompt that has worked well for me:
In the repository, the function that parses dates in the import module accepts
both ISO and US formats but the US branch drops the timezone. Fix the US branch
so it preserves the timezone, add a unit test that fails before the fix and
passes after, and do not touch any other file. Summarize the diff when done.
What matters: one file's worth of change, an explicit test, an explicit "do not touch" boundary, and a request for a summary so the review starts from the agent's own description.
The app announcement describes developers "delegating work, running tasks in parallel". On the web that means opening several tasks rather than one large one. Split by independence, not by size:
Task A: add input validation to the signup endpoint; tests included.
Task B: replace the deprecated logging calls in the worker package; no behavior change.
Task C: write the README section for the new CLI flag; documentation only.
Each task should be mergeable on its own. If B depends on A, run them sequentially instead; a parallel run that conflicts costs more review time than it saves. Remember that rate limits are shared across the web, app, CLI and IDE, so three parallel cloud tasks draw from the same pool as a local session.
The announcement talks about agents handling tasks that "span hours, days, or weeks". For anything you will not be watching, write the prompt like a handover note:
Goal: migrate the test suite from the old assertion library to the built-in one.
Constraints: keep every test's intent; do not delete tests; keep CI green.
Verification: the full test command must pass at the end.
If a test cannot be migrated mechanically, leave it unchanged and list it in
the final summary with the reason.
The last line is the important one. Giving the agent a legitimate way to stop short prevents it from forcing a change to satisfy the goal.
Before a batch of tasks, open the Codex pricing page and confirm what your plan allows today. The doubled limits and the Free and Go inclusion were announced as a limited-time change, and the pages used for this repository do not give the numbers, so the pricing page is the only reliable source. Enterprise teams should read the enterprise page instead.
Codex web earns its keep on repositories: multi-file changes, tests, pull requests. When the request is "I need a page" or "clone this site's layout for a prototype", the repository workflow is overhead. Try Begin.sh - turn a prompt or a URL into a downloadable static site or Expo app. You get a zip, host it anywhere, and there is no hosting, backend or auth to configure or pay for. Use the agent for code that lives; use the generator for artifacts you ship once.
Last reviewed: 2026-09-22