Reference · OpenAI · updated 27 September 2026
The cleanest token story of the group: short-lived installation tokens per operation. Here is what its public documentation says about authentication, commit attribution, the sandbox it runs in, how you hand it work, and the limits you will hit.
| Auth | A GitHub App (chatgpt-codex-connector). OpenAI states it uses short-lived, least-privilege GitHub App installation tokens for each operation and respects branch protection. |
| Attribution | Reviews are posted by the connector bot. The docs do not state who the commit author is on Codex-authored changes. |
| Sandbox | OpenAI-managed containers in two phases: a setup phase with network access, then an agent phase that is offline by default. |
| Task intake | The ChatGPT interface, @codex or @codex review on a PR, GitLab merge requests, Linear, Slack, and a Codex GitHub Action. |
| Limits | Bound to OpenAI usage plans. The sandbox is OpenAI-hosted only; there is no self-hosted execution target. |
Unverified: Commit author identity comes from third-party repositories, not vendor documentation.
Every cloud agent in this series invents its own identity and attribution model, because the host gives them nothing to stand on. The consequences are the same across vendors: agents share a person's token or an org-wide app installation, so a fleet hits secondary rate limits as one user; attribution lives in free-text trailers the host never reads; and rules like "a human must approve changes under this path" are left to the agent to enforce on itself.
Paste this to your coding agent. A human approves one code on github.com; the agent does the rest.
Set up githug for owner/repo: run `curl -fsSL https://githug.ai/install | sh`, then `hug init --repo owner/repo --json`. Show me any code or link it prints and wait for it to finish.
Quickstart · Set it up yourself · free during the beta