Dev pairs Codex with a local GPU model as a safe second opinion

Someone put an already-owned RTX 3090 (24GB GPU) to use in a hybrid setup meant to stretch usage on Codex. Codex (running the model) stays the main agent and decision-maker at all times.

It sends a bounded code snippet, diff, log excerpt, or question to a , , running through Ollama on the local GPU. The returns a read-only second opinion, but Codex independently verifies whatever comes back and decides what, if anything, to actually use.

The connection runs through a local MCP server called local_gpu_reviewer, and that reviewer is deliberately given no -write, shell, Git, commit, deployment, publishing, or email tools — it can only analyze text explicitly handed to it by Codex. The trickiest technical snag was that Codex's native custom-agent route sent messages in a format the setup didn't initially expect.

Key points

  • Repurposes an owned RTX 3090 (24GB GPU) as a er
  • Codex remains the primary agent; the ( via Ollama) only offers a secondary opinion
  • only receives bounded text (snippets, diffs, logs) and has no execution or deployment access
  • Connected via an MCP server bridging Codex and the
  • Codex always independently verifies the 's findings before acting
Read original