Same Codex desktop — diffs, review mode, browser annotations, Git worktrees. Different brain. Ollama’s Codex App integration lets you keep OpenAI’s coding UI while inference runs on a local model or Ollama Cloud. One command wires it:
ollama launch codex-appThat’s the whole pitch. Not a new agent, not a fork of the IDE — a provider swap with the workspace features intact. For anyone already living in Codex on macOS or Windows, this is how you stop choosing between “good desktop agent” and “model stays on my box.”
What stays, what moves
Codex still owns the repository, worktrees, command execution, and review loop. Ollama serves models through its OpenAI-compatible endpoint (typically localhost:11434). Local weights stay on the machine; Ollama Cloud models show up when you want something too big for the laptop. Coverage from Zeniteq spells out the split cleanly: Ollama handles inference; Codex keeps the coding surface.
Optional flags matter in practice. Pin a model on launch (ollama launch codex-app --model gemma4:31b or a cloud id), and --restore flips Codex back to the previous profile when you need OpenAI’s defaults again. Ollama backs up config under ~/.ollama/backup/codex-app/ before it overwrites settings — small detail, big for people who live in that app all day.
Why I care (and why it’s not “just another chat”)
I already bounce between Cursor Projects, Codex, and local stacks. The win here is not abandoning the desktop agent when a repo shouldn’t leave the LAN. Pull a coding model, launch, select it in Codex, keep using review comments and isolated worktrees. Privacy is bounded — Codex can still open browsers, hit package registries, and talk to local servers — but the model prompt path can stay local. That distinction is the whole point for security-minded workflows and client work that never belongs on a hosted inference endpoint.
Manual fallback exists if you want profiles: point ~/.codex/config.toml at Ollama’s /v1 endpoint with the responses wire API. Most days the launcher is enough. When a model change needs a restart, let Ollama bounce Codex rather than chasing stale config by hand.
What this isn’t
It’s not “every ChatGPT plugin now runs on Llama.” Docs and reporting both emphasize Codex mode features — browser, review, worktrees — not blanket desktop-plugin parity. Treat it as a coding-agent provider swap. Model quality and tool reliability still decide whether a multi-file refactor lands cleanly; the integration only makes the comparison fair inside one UI.
For anyone already deep on Ollama plus desktop agents: this closes a gap. Same Codex muscle memory, optional local brain, restore when you need the hosted path. That combination is now a first-class workflow, not a weekend hack — and it sits cleanly beside Cursor without forcing you to abandon either.
Sources: Ollama Codex App docs · Zeniteq on Codex + Ollama