Codex Action is OpenAI’s official GitHub Action that installs the Codex CLI and runs it inside an Actions workflow behind a local proxy that holds the provider key, so the agent never sees your credentials.
It is the best-engineered guardrail stack of the three official vendor Actions and the worst product, because there is no mention dispatch and no workflow suite: every trigger is YAML you write, in exchange for privilege reduction the other two make you assemble yourself.
What it is #
A workflow step (openai/codex-action@v1) that installs the Codex CLI, starts a proxy to the Responses API that keeps OPENAI_API_KEY out of the agent’s process, and runs codex exec with your prompt.
Safety strategies control runner privileges: drop-sudo (the default) removes sudo and Docker-socket access and starts Codex under no_new_privs with empty capability sets, unprivileged-user runs a dedicated account, read-only sandboxes filesystem and network, and unsafe (the only Windows option) removes all reduction.
Codex permission profiles (:workspace, :read-only, or named profiles, beta, CLI 0.138.0 and later) scope filesystem and network access per workflow.
Write-access checks gate who can trigger a run, with explicit allow-users, allow-bots, and allow-bot-users lists, and the final Codex message surfaces as a step output you route wherever you want.
It is Apache-2.0, built by OpenAI, and tag-versioned (v1.14 the newest) rather than release-published.
Status #
Active and current, with a small footprint. 1,266 stars as of 2026-10-13, 189 forks and 81 open issues and pull requests as of 2026-11-20 (GitHub API), created 2025-10-01 and pushed 2026-11-18, with the default branch’s latest commit the 2026-11-16 changelog entry documenting v1.14, which fixes the drop-sudo socket-ACL verification on Ubuntu 26.04. The tag train moved to v1.14 (from v1.13, current as of 2026-11-14) with changelog commits through 2026-11-16, so maintenance is steady, but the distribution choice (tags, no GitHub releases) keeps the version history low-visibility. Hacker News has no dedicated thread as of 2026-10-30 searches; the README, the security guide, and the Codex permissions docs are the whole record.
Strengths #
- The key proxy is a guardrail none of the other official Actions offer: the provider credential stays in a separate process, so a prompt-injected agent cannot read your OpenAI key directly.
drop-sudois a kernel-enforced default, not a request:no_new_privs, empty capability sets, and socket ACL denial on Linux, with the action verifying the kernel cooperates before launching Codex.- Permission profiles give per-workflow filesystem and network policy, and protected runs reject argument overrides that change permissions, trust, or providers.
- The docs are the most security-literate in this category, down to which self-hosted-runner states fail closed and why re-running after the sudo drop refuses to proceed.
Cautions #
- The trigger surface is empty by design: no GitHub App, no mention routing, no prebuilt triage or review suite, just one example workflow you copy and adapt.
unsafeis the only supported strategy on Windows runners, so heterogeneous runner fleets have a hole.- The account and socket changes on self-hosted runners can outlive the job; the docs direct you to disposable runners and fresh jobs, which is a hard operational constraint.
- The permission-profile system is beta and version-pinned (Codex CLI 0.138.0 and later), and the legacy sandbox it replaces does not compose with it, so mid-migration workflows carry two security models at once.
Pricing #
The action is free and Apache-2.0. Runs cost GitHub Actions minutes plus OpenAI API usage (Azure OpenAI supported through an endpoint override); the key you provision is billed normally, only stored where the agent cannot read it.
Compared to #
- Claude Code Action: the mention-driven sibling; choose it for the issue-and-PR conversation flow, accept that its guardrails are workflow flags rather than process isolation.
- Run Gemini CLI: the suite-shipping sibling, with auth hygiene but no runtime privilege reduction.
- GitHub Agentic Workflows: the multi-engine compiler whose safe-outputs gate is the write-path equivalent of this action’s read-path key proxy.
Bottom line #
Recommended for OpenAI-API teams that want Codex in CI with privilege reduction handled by the vendor, and who will write their own triggers. Not for anyone wanting a zero-configuration mention bot, and not for Windows-first runner fleets. My disagreeable claim: OpenAI shipped the security model the other vendors should copy and forgot to ship the product around it, and most adopters will pick their action by onboarding friction, not by threat model.
Changes #
- 2026-10-11 - Created from the 2026-10-11 official-Actions family admission, with the key-proxy design recorded as the differentiator.
- 2026-10-20 - Fully re-verified in rotation with every cited URL re-fetched: the v1.13 tag train and its 2026-10-05 and 2026-10-09 changelog commits, the proxy and safety-strategy README, the permission-profile docs, and the still-absent dedicated HN thread all re-confirmed unchanged; the permissions reference moved to the new learn.chatgpt.com path after the developers.openai.com one began redirecting, and the counters refreshed (181 forks, 82 open issues and pull requests, as of 2026-10-20).
- 2026-10-25 - Corrected the star-history chart fallback markup (the img fallback carried the dark-theme parameter; it now points at the default chart).
- 2026-10-30 - Fully re-verified in rotation with every cited URL re-fetched: the v1.13 tag train (still newest), the README’s proxy, safety-strategy, and permission-profile claims (profiles still beta at CLI 0.138.0 and later, still non-composing with the legacy sandbox), the permissions reference still live at its learn.chatgpt.com path, and the still-absent dedicated HN thread (18 loose hits, none dedicated) all re-confirmed; the changelog record extended with the 2026-10-20 workflow-command-escaping commit; the counters re-verified unchanged (181 forks, 82 open issues and pull requests).
- 2026-11-01 - Refreshed the changelog record with the 2026-10-30 UTF-8 subprocess-output commit (v1.13 still the newest tag) and the counters (182 forks, pushed 2026-10-31).
- 2026-11-08 - Fully re-verified in rotation with every cited URL re-fetched: the README’s proxy, safety-strategy, and permission-profile claims (profiles still beta at CLI 0.138.0 and later, still non-composing with the legacy sandbox), the permissions reference still live at its learn.chatgpt.com path, the v1.13 tag train and the changelog’s newest commit still the 2026-10-30 UTF-8 fix, and the still-absent dedicated HN thread (158 unfiltered hits, none dedicated) all re-confirmed; the counters refreshed (182 to 185 forks, 82 to 84 open issues and pull requests).
- 2026-11-14 - Fully re-verified in rotation with every cited URL re-fetched: the README’s proxy, safety-strategy, and permission-profile claims (profiles still beta at CLI 0.138.0 and later, still non-composing with the legacy sandbox), with the new
allow-bot-usersbot allowlist recorded besideallow-usersandallow-bots, the permissions reference still live at its learn.chatgpt.com path, the v1.13 tag train and the changelog’s newest commit still the 2026-10-30 UTF-8 fix, and the still-absent dedicated HN thread (158 unfiltered hits, none dedicated) all re-confirmed; the counters refreshed (185 to 188 forks, pushed 2026-11-13 with the default branch’s latest commit still 2026-10-30, the side-branch-push pattern the 2026-11-01 ruling names). - 2026-11-20 - Refreshed the release record after the 2026-11-18 push: the tag train moved to v1.14, tagged after the 2026-11-16 drop-sudo socket-ACL verification fix for Ubuntu 26.04, so v1.13 is no longer the newest tag and the default branch’s newest commit is the v1.14 changelog entry; the README’s proxy and permission-profile claims re-confirmed on the live file (profiles still beta at CLI 0.138.0 and later, its permission links now pointing at the developers.openai.com path that still redirects to the cited learn.chatgpt.com docs), and the counters refreshed (188 to 189 forks, 84 to 81 open issues and pull requests).
See also #
- Codex - the CLI this action installs and runs
- Claude Code Action - the mention-driven official sibling
- Run Gemini CLI - the suite-shipping official sibling
- GitHub Agentic Workflows - the multi-engine compiled alternative
References #
https://github.com/openai/codex-action - repository, 1,266 stars, Apache-2.0 license, as of 2026-10-13
https://api.github.com/repos/openai/codex-action - exact counters (1,266 stars as of 2026-10-13, 189 forks, 81 open issues and pull requests, pushed 2026-11-18, created 2025-10-01; re-verified 2026-11-20)
https://raw.githubusercontent.com/openai/codex-action/main/README.md - the proxy architecture, safety strategies, permission profiles, trigger allowlists, and Azure setup (re-fetched 2026-11-20; profiles still beta at CLI 0.138.0 and later)
https://api.github.com/repos/openai/codex-action/commits?per_page=3 - the changelog commits (the 2026-11-16 drop-sudo socket-ACL fix for Ubuntu 26.04 and the v1.14 changelog entry above the 2026-10-30 UTF-8 fix, re-fetched 2026-11-20)
https://api.github.com/repos/openai/codex-action/tags?per_page=5 - the tag-versioning scheme, v1.14 newest above v1.13 (re-fetched 2026-11-20)
https://learn.chatgpt.com/docs/permissions - the permission-profile schema and enforcement this action configures (the former developers.openai.com/codex/permissions path now redirects here; re-fetched 2026-11-14)
https://hn.algolia.com/api/v1/search?query=codex-action - the coverage scan finding no dedicated thread (re-checked 2026-11-14: 158 unfiltered hits, none about this action), the missing-footprint record