Aident AI

Codex App Update Requires macOS 13? Recover on macOS 12
If a Codex or ChatGPT desktop update now says it requires macOS 13 or later, stop retrying the installer. Reports on August 5, 2026 show ChatGPT build 6223, version 26.730.61309, declaring macOS 13 as its minimum and failing on macOS Monterey 12.7.6. OpenAI's current requirements are stricter: the combined ChatGPT app that includes Codex officially requires macOS 14, and OpenAI says it will not release that app for older operating systems.
The durable choices are to upgrade macOS on supported hardware, use the Codex CLI on macOS 12, or use Codex on the web. A rollback should be temporary and should use only an OpenAI-hosted, correctly signed build that matches your Mac's architecture. Do not install an unverified mirror or disable Gatekeeper to make an old package run.
Confirm the Operating System, Architecture, and App Boundary
Run these read-only checks in Terminal:
Expected result for the reported failure:
sw_versstarts with12..uname -mreturnsx86_64on an Intel Mac orarm64on Apple silicon.The updated app reports build 6223 or another build whose minimum system version exceeds your installed macOS version.
Launching it returns an incompatibility dialog or
NSOSStatusErrorDomain Code=-10825.
If macOS already satisfies the app's declared minimum, this is not the Monterey compatibility failure. Check the code signature and app logs instead of downgrading.
Why Reinstalling the Same Update Cannot Work
A macOS executable records the oldest operating system it supports. The August 5 report for build 6223 found LC_BUILD_VERSION minos 13.0. macOS 12 rejects that binary before the application can migrate state, connect to Codex, or display its normal interface.
Changing permissions, clearing caches, or launching with sudo cannot lower the binary's minimum OS. Removing quarantine or ad-hoc signing the app only weakens its security checks. The package must support the installed OS, or the OS must be upgraded.
There is also a product boundary to account for. Since July 9, the new ChatGPT desktop app combines Chat, Work, and Codex. OpenAI's current help page lists macOS 14 as the supported minimum for that combined app. A prior build that happens to start on macOS 12 is therefore a short compatibility bridge, not a supported long-term release channel.
Choose the Safest Recovery Path
Your situation | Recommended path | Tradeoff |
|---|---|---|
Your Mac supports macOS 14 | Back up your work, update macOS, then install the current app | Restores the supported desktop path but requires an OS upgrade |
You must remain on macOS 12 | Use Codex CLI or Codex on the web | Keeps Codex available without the combined desktop app |
You need the old desktop app briefly | Restore an OpenAI-hosted signed build after verifying it | Temporary and potentially behind on fixes |
You only found a community mirror | Do not install it | The source and bytes are not established by the issue report |
Do not erase ~/Library/Application Support, ~/.codex, or project worktrees as part of this recovery. The launch failure is an OS compatibility boundary, not evidence that your sessions or repositories are corrupt. If chats appear missing after an update, use the inventory-first steps in Recover Codex Chats After an Update.
Option 1: Use the Supported Codex CLI on macOS 12
The Codex CLI documents macOS 12 or later as its minimum. OpenAI's current standalone installer selects the correct native binary:
Expected result: codex --version prints a CLI version, and codex opens the terminal interface and offers ChatGPT sign-in.
If you prefer a manual release asset, match the architecture exactly:
aarch64-apple-darwinforuname -moutputarm64.x86_64-apple-darwinforuname -moutputx86_64.
The official Codex repository links both assets. Do not assume a 2020 Mac is Intel: Apple shipped both Intel and M1 models that year.
You can also use Codex at chatgpt.com/codex while deciding whether to upgrade the operating system.
Option 2: Verify an Official Previous Build Before Rollback
An August 5 issue comment recovered an older x86_64 package URL from the local updater history under persistent.oaistatic.com. The same thread also shows why copying a mirror link is unsafe and why architecture matters: an M1 user first downloaded the Intel build and only recovered after selecting the correct architecture.
Search your own updater caches and logs for an OpenAI-hosted URL:
Only continue if the exact URL starts with https://persistent.oaistatic.com/. If you cannot recover an official URL for the correct architecture, use the CLI or web path instead of guessing a filename or using a third-party mirror.
Download the recovered URL without opening it:
Extract it into a new temporary directory:
Now verify compatibility and provenance before moving anything:
Continue only when all of these are true:
The candidate's
LSMinimumSystemVersionis no newer than your installed macOS.Its executable contains your native architecture.
codesignreports that the app is valid on disk and satisfies its designated requirement.spctlaccepts it under the Mac's current policy.The candidate and your previously installed official app show the same signing authority and team identifier.
Apple notes that codesign verifies sealed contents and internal consistency, while spctl evaluates the Mac's current Gatekeeper policy. Neither check makes an arbitrary mirror trustworthy, which is why the asset host check comes first.
Back Up the Current App and Test the Candidate
Quit the app, move the incompatible bundle to a recoverable location, and copy in the verified candidate:
Expected result: the app opens without the minimum-OS dialog. Confirm its version in the About dialog, then test one existing project without deleting or migrating any local state.
If the copy fails or the candidate does not launch, remove only the failed candidate and restore the backup:
Do not disable Gatekeeper, clear quarantine attributes, or re-sign the app. Do not use undocumented preference writes as a permanent way to suppress updates. A rolled-back build can miss security and reliability fixes, so move to the CLI, web, or a supported OS and retire it promptly.
Common Failure Modes
The old build still says it is incompatible
Check LSMinimumSystemVersion. A version number alone does not prove Monterey support, and the archive may contain a different product generation than the one you expected.
The package is x86_64 but the Mac is arm64
Recover the official arm64 asset instead. Rosetta can run many Intel applications, but it does not make an Intel-only rollback the correct or lowest-risk package for Apple silicon.
Gatekeeper rejects the candidate
Stop. Redownload it from the exact OpenAI-hosted URL and repeat codesign and spctl. Do not solve a failed signature check with xattr, ad-hoc signing, or a mirror.
The app updates into the incompatible build again
Move to the CLI or web immediately. OpenAI's current requirements make clear that old operating systems will not receive a compatible combined app, so repeatedly pinning an old desktop build is not a durable fix.
Keep Connected Work Available From the CLI
If the desktop app was your path to external tools, Aident Loadout can keep authentication and Action discovery outside the app bundle. From the Codex CLI, install or update Aident by asking your agent to follow https://aident.ai/SETUP.md, then verify one read-only path:
Inspect and preflight the exact Action before execution. Expected result: the CLI reports authenticated account and Vault state, then returns a reviewed Action contract without exposing provider credentials.
Ready to confirm the fallback works? Set up Aident Loadout and preflight one read-only Action.
Sources
OpenAI Codex issue 37002: Unable to install after clicking Update
OpenAI Codex issue 37019: Auto-update breaks macOS Monterey support
Review this article when OpenAI changes the combined app's macOS requirements, issues 37002 or 37019 close with an official recovery path, Codex CLI drops macOS 12, or the documented Aident Loadout setup flow changes.



The one tool
for every tool
your agent needs.
Give any AI agent real capabilities in seconds. Connect 1,000+ tools once, skip the setup headache, and let your agents execute.