Facts checked Nov 11, 2026 · all comparisons
tiny vs Codex cloud
Codex cloud runs Codex tasks in containers on OpenAI's infrastructure. You start a task against a repository from your connected GitHub or GitLab account, Codex works on it in its own workspace while your computer sleeps, and you review the result.
tiny runs the open-source Codex CLI, unmodified, as a session on a Kubernetes cluster you own, next to Claude Code sessions in the same fleet.
Side by side
| Codex cloud | tiny | |
|---|---|---|
| where it runs | OpenAI-managed containers | pods on your Kubernetes |
| how you work with it | start a task, review the result | an interactive session you can attach to mid-task, or a task delivered from an issue |
| internet access | configurable: everything, package-manager domains, or an allow-list | default-deny NetworkPolicy; optional hostname allow-list proxy |
| repositories | your connected GitHub or GitLab account | any repo the session can clone; for a private one, a deploy key, or ship a local folder with --dir |
| git credentials in the agent's environment | not described in OpenAI's docs | none: work leaves as a git bundle a separate job pushes |
| agents | Codex | Codex and Claude Code |
| what you operate | nothing | a cluster |
Where Codex cloud is the better choice
- Nothing to run. No cluster, no images, no upgrades.
- Fire-and-forget tasks that you start from ChatGPT and review later are exactly what it's built for.
Where tiny fits
- Your infrastructure. The code, the workspace and the session transcript stay on your cluster. The model API still receives prompts, as with any hosted model.
- Interactive. Attaching puts you in the actual Codex terminal, so you can steer a task that's going the wrong way.
- One fleet for both agents. Codex and Claude Code sessions share the fleet screen, the human gate, the egress rules and the GitHub loop.
- No credentials in the pod by default, with every exception listed in the threat model.
Sources: OpenAI's Codex cloud documentation, checked on the date above. Where OpenAI's docs don't say, the table says so rather than guessing. Tell us on GitHub if a row has gone stale.