Codex CLI 0.169.0, published on November 13, 2026, wins this comparison for any non-interactive use. In 0.169.0, codex exec keeps project configuration and MCP servers disabled in untrusted directories, even with --sandbox workspace-write. Builds before it would still start the MCP servers a freshly cloned repository declared. If you run Codex in CI, scripts or cron, upgrade, then decide per job whether to pass --trust or move the servers you need into your own user configuration.
TL;DR
- Codex CLI 0.169.0 changes
codex execso that repository-defined project configuration and MCP servers no longer load in untrusted directories. - The behaviour holds even with
--sandbox workspace-write, which previously did not stop a cloned repo from choosing which helper commands start. -
--trustis new on bothcodex execandcodex review. It saves project trust for the current and future chats. - The change is in PR #60041, listed first under Bug Fixes in the 0.169.0 release notes.
- CI jobs that relied on a repo's own MCP servers will lose them after the upgrade. Audit
.codex/before granting trust.
What changes for codex exec untrusted directories in 0.169.0?
The release notes put it in one sentence: codex exec keeps project configuration and MCP servers disabled in untrusted directories, and --trust on codex exec or codex review explicitly saves project trust. The pull request adds a warning when project configuration is skipped, so a job that silently depended on repo config now says so in its log.
| Situation | Before 0.169.0 | 0.169.0 |
|---|---|---|
codex exec in a fresh clone, no flags |
Project config and repo MCP servers could load | Both stay disabled, with a warning |
codex exec --sandbox workspace-write |
Still loaded them | Still disabled |
codex exec --trust |
Flag not available | Saves trust for the current and future chats |
codex review |
Same exposure as exec
|
Same fix, and --trust is available |
--trust with --ignore-user-config
|
Not applicable | Refused |
--trust on an explicitly untrusted project |
Not applicable | Refused |
The last two rows come from the PR description: the trust grant refuses explicitly untrusted projects and cannot be combined with --ignore-user-config. The diff is +193 and -69 across 7 files, merged on November 7, 2026, so it is a contained change rather than a rewrite of the trust model.
Why was loading repo MCP servers in codex exec a problem?
An MCP server entry in a configuration file is a command. When the agent starts, Codex launches it. If that file lives in the repository, then whoever controls the repository chooses what runs on your machine at session start.
In an interactive session you at least get a moment to see a trust prompt. A codex exec run in CI has nobody at the keyboard. The sandbox setting does not help much here, because the PR notes that trust applies to MCP commands outside the sandbox. A pull request from a fork, a dependency checked out for analysis, or a repo cloned by a cron job could all carry a project config that names a server you never reviewed.
That is the clone-and-run pattern familiar from build scripts and postinstall hooks. The agent version of it is worth taking seriously, and it sits in the same family of risks we covered in the Cursor breach and Claude Code backdoor write-up: tools that hold credentials and launch helper processes are attractive targets.
How do I migrate a CI job to Codex CLI 0.169.0?
Work through these steps for each job that calls codex exec or codex review.
- Pin the version to 0.169.0 or later, so behaviour does not differ between runners.
- Run the job once and read the log. Expect a warning wherever project configuration was skipped.
- List which MCP servers the job actually needed. If none, you are finished.
- For servers you own and maintain, move the definition into user-level configuration on the runner, outside the repository. Then the repo cannot change it.
- For servers that must stay in the repo, review
.codex/in the exact commit being built, then pass--trust. Do this only for repositories your organisation controls. - Never combine
--trustwith checkouts of untrusted pull requests. Run those jobs without it and accept the reduced configuration.
A minimal before and after for a trusted internal repo:
# Before: relied on the repo's MCP servers loading by default
codex exec --sandbox workspace-write "run the migration checks"
# After: trust is explicit, and only for a repo you have reviewed
codex exec --trust --sandbox workspace-write "run the migration checks"
One tradeoff to be honest about: --trust saves trust for future chats too. On a persistent runner that means later runs in the same directory inherit it. On ephemeral runners the effect disappears with the container, which is the safer arrangement.
Is 0.169.0 better than 0.168.0 for interactive use?
Mostly the same, with different features. Both versions landed on November 13. The 0.169.0 release, which the release page timestamps at 08:37 UTC on November 13, adds multiple side-chat tabs in the opt-in terminal workspace, task search in the agent command center, terminal setup for custom model providers, and a device-code option for supported gateways over SSH. Codex CLI 0.168.0 also arrived that same day, with the tui_workspace side conversation and a model-provider browser summarised by AI/TLDR, and AgenticMart's November tracking table shows peer tools shipping alongside it: OpenCode 2.3.0, GitHub Copilot CLI 1.0.112 and goose 1.59.1. These are conveniences. The trust change is the item that should alter your operations.
If you only use Codex at a terminal in repositories you wrote, you will notice little. If you automate it, the verdict is simple: 0.169.0 is the version to run.
What does Codex cost in India?
Codex usage comes with ChatGPT plans, so the trust change does not alter pricing. ChatGPT Plus is Rs 1,999 per month in India, inclusive of 18% GST, and it includes expanded Codex usage. ChatGPT Go is Rs 399 per month, also with GST included, after its promotional period ended on November 7, 2026. Codex run with an API key follows API pricing instead of the subscription, which matters for CI where you would normally use a key.
How does this fit with other coding agents?
Trust prompts for project-defined tooling are becoming a baseline expectation. For a peer release, see our notes on Claude Code 2.1.323 and code mode. For how open-source agents handle delegated work, see OpenCode background subagents versus inline. On rollout practice, verify-first versus build-first agentic engineering is a useful frame for deciding when a job earns --trust.
FAQ
Q: Does codex exec still load MCP servers from my user configuration?
A: The change targets project configuration and the MCP servers a repository declares in untrusted directories. User-level configuration is not what the fix disables, which is why moving needed servers there is a sensible migration path. Check your own logs after upgrading to confirm.
Q: Does --sandbox workspace-write make a repo trusted?
A: No. As of 0.169.0, codex exec keeps project configuration and MCP servers disabled in untrusted directories even with --sandbox workspace-write. Sandbox scope and project trust are separate decisions.
Q: How do I grant trust in a non-interactive run?
A: Pass --trust to codex exec or codex review. It saves project trust for the current and future chats. It refuses explicitly untrusted projects and cannot be combined with --ignore-user-config.
Q: Does this affect codex review as well?
A: Yes. The same fix line covers both commands, and --trust works on codex review.
Q: Which pull request made this change?
A: #60041, titled "Require explicit project trust in non-interactive runs", merged on November 7, 2026 and shipped in 0.169.0.
Q: Should I pass --trust everywhere to keep old behaviour?
A: No. That recreates the exposure the release removed. Use it only on repositories you control, after reviewing .codex/ in the commit being built.
Top comments (0)