Codex could optionally use the user’s own hardware for suitable low-risk work — repository indexing/retrieval, search, simple edits, linting/tests, and potentially a small local model — while keeping frontier reasoning and uncertain/high-risk decisions in the cloud.
The local layer should never be able to silently hide relevant context from the cloud model; the cloud model must always be able to request more source context or escalate a task.
Make it opt-in, automatically benchmark available hardware, and allow configurable CPU/RAM/VRAM limits.
Potential benefits:
- lower OpenAI inference load and cost;
- higher effective Codex usage limits;
- less repository data needing to leave the machine;
- better use of powerful consumer/workstation hardware;
- cloud frontier intelligence remains available whenever it matters.
This could range from local search/indexing only on modest machines to substantially more local agent execution on systems with capable GPUs.