Here's a tip for anyone whose Codex usage limit fills up quickly.
Telling it to make direct use of your own computer's hardware can have the same effect as a 5x bigger limit.
Codex Called llama.cpp
This tip was posted on X by AI developer Mike Bradley at 12:12 AM (KST) on August 18, 2026.
Original: "Just tell Codex to aggressively use your hardware. With a bit of setup, it's the equivalent of increasing your usage limits by 5x."
The post also got replies saying people had reproduced it.
Developer Johnny Everson replied, "I told Codex to use OMP as an implementer subagent, had OMP point to a llama.cpp model running locally, and it just worked."
Why Does the Limit Seem to Stretch?
Codex's cloud token usage is billed per task.
If you hand some tasks to local hardware (your own GPU, a local LLM, and so on), the cloud side burns that many fewer tokens.
In a follow-up reply, Bradley added, "It's definitely worth it even for people who max out the Max plan."
But the specific setup steps and how the effect differs by task weren't detailed in the thread.
Hand Off Repetitive Work First
If you often hit Codex usage limits, it's worth considering splitting repetitive, heavy subtasks (builds, tests, local model inference, and so on) off to a local subagent.
But this isn't a feature OpenAI has documented; it's a community workaround users trigger through prompts.
"5x" Isn't an Official Figure
The "5x" figure is Bradley's personal estimate from experience, not a value verified by OpenAI.
The effect can vary widely with your hardware specs and the type of work.
It also hasn't been officially confirmed whether this approach is fine under Codex's terms of service.