Connecting your own ChatGPT/Codex subscription: what you're agreeing to

What we actually hold, and where it runs

When you connect your subscription, Code Voucher stores your OpenAI credentials, encrypted, in our database. That's expected of any "connect your account" integration.

During a review or patch job, a trusted token broker supplies a short-lived access token to the Codex process over its private input stream. Refresh tokens stay encrypted in the database and are used only by the trusted broker. Jobs do not receive a credential file or a refresh token, and cannot rotate a shared refresh token independently.

The Codex process still reads repository content while authenticated to your account. Untrusted instructions in that content can try to manipulate its behavior, consume your quota, or influence its output. Sandbox restrictions reduce that exposure, but an access token remains a credential and the runtime must still be treated as sensitive. Reconnecting replaces the credential generation; existing jobs cannot update or invalidate that new connection.

Ways your credentials or quota could still leak or be misused

Even with our safeguards in place, there are residual paths a malicious or compromised repository could use:

• Content Codex writes into a proposed fix (a "patch merge request"). Automatic proposed-fix publishing isn't turned on for subscription connections yet, but it's a planned capability and the underlying risk is the same as for review: if a run produces a suggested code change, and something in your repository has manipulated the model into embedding sensitive text there, that text could ride along into what gets proposed. We're calling it out now so it isn't a surprise later.
• Rich summaries of what the review found. If you opt in to more detailed, model-written summaries (instead of the private dashboard-only report), that summary is also model-generated text, and it inherits the same risk: something in the repository could try to smuggle unexpected content into it.
• Compromise of the authenticated process. The temporary Codex process holds a short-lived access token in memory. A runtime compromise could expose that token, even though no refresh-token file is placed in the workspace.

None of this means a leak is likely on every run. It means a subscription connection removes a layer of protection that API-key-based, credential-free review does not have to give up, and you should treat that trade-off as real, not theoretical.

Publication stays private by default

By default, review results are only visible to you on the Code Voucher dashboard — nothing is posted back to your repository host automatically. If you want review summaries posted as merge request comments, that's a separate, explicit setting you opt into on top of the subscription connection. Connecting your subscription does not, by itself, turn on any automatic posting.

This is your OpenAI account, and your responsibility

Using a personal or organizational ChatGPT subscription to run an automated tool like Codex may not be something OpenAI's terms of service clearly contemplate. We aren't able to tell you definitively whether this complies with your OpenAI plan's terms — that determination, and the standing of your OpenAI account, is your responsibility as the account holder, not Code Voucher's.

Our recommendation: use a dedicated, low-value account

Because of the above, we recommend connecting a ChatGPT account that:

• Isn't your personal or primary work account,
• Isn't shared with other sensitive usage, and
• You'd be comfortable resetting or losing access to if something went wrong.

Treat it the way you'd treat any credential handed to an automated pipeline that processes third-party or untrusted input: useful, but replaceable.

Disconnecting

You can disconnect your subscription at any time from the settings page. Disconnecting removes the stored credentials from Code Voucher immediately.

Disconnecting from Code Voucher is not the same as revoking the connection on OpenAI's side. To fully invalidate the credential (so it can no longer be used by anyone who may have obtained a copy of it), you also need to go to your OpenAI/ChatGPT account settings and revoke the session or sign out of that connected app there. We recommend doing this as a matter of course when you disconnect, and immediately if you have any reason to suspect the credential was exposed.

Questions

If anything here is unclear, please reach out before connecting — this is meant to be a real decision, not a formality.