Codex Desktop macOS consumes token quota while idle before any prompt, even after state reset and MCP disabled
What version of the Codex App are you using (From “About Codex” dialog)?
26.422.30944 (2080)
What subscription do you have?
Business
What platform is your computer?
Darwin 25.4.0 arm64 arm
What issue are you seeing?
Tokens are being consumed before any prompt. Started the workday, haven't done anything yet, not automations configured. At some point a software-controlled instance of Chrome opens and starts navigating my local project dirs and files. I kill Codex and it stops. I start it again and it starts navigating again. I check the 5h token limit and I have 88% remaining (no prompt today yet).
I disable the Chrome MCP. I ask ChatGPT, among other things, it suggests running:
rm -rf ~/.codex
rm -rf ~/Library/Application\ Support/CodexAfter I log back in (lost all chat history) remaining 5h tokens is down to 84%. No MCP or anything configured. A few minutes pass, I check again - down to 78%.
I don't know what else I can try.
What steps can reproduce the bug?
I just opened a fresh instance of the Mac Codex app and logged in. Didn't do anything special.
What is the expected behavior?
No token usage when idle.
Additional information
_No response_
3 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
fwiw, I have a suspicion for what caused this:
<img width="705" height="91" alt="Image" src="https://github.com/user-attachments/assets/9863f757-16f1-4d27-a5c8-7dc21d1eb8f0" />
(it was enabled and I disabled it now - not sure if the token consumption stopped before I disabled it or due to the disabling)
If this is indeed the cause of this, I see it as a product failure. If a background process consumes tokens and uses tools that have visible side-effects, the app UI should make it clear that this is happening (a spinner with a tooltip would do it) and ideally an option to stop it, if the user chooses to. The toggle in settings should also make it clear that this will use tokens. Perhaps let users control which model does this job or even put a cap on token usage? Maybe it should also optimize for the 5h token limit by using up tokens only before the reset.
I have the same issue they are stealing tokens