Ollama 0.24 ships the Codex App, OpenAI's desktop experience for running Codex threads in parallel. It lands with built-in worktree support and git functionality baked in. For builders juggling multiple feature branches or experimental agents, that means less context switching across terminals and editors.
Launch it with a single command:
ollama launch codex-app
The desktop experience has three features worth paying attention to.
Built-in browser. Codex can load local servers and sites directly inside the app. You annotate on the page and request changes from there. No more screenshotting your localhost UI, pasting it somewhere, and describing what you want. You point, you annotate, you iterate.
Review mode. You can read code inside the app, leave comments, and loop back without leaving your workspace. This closes the gap between generation and review in a single surface.
Model choice. The release ships with a short, opinionated list of models. For difficult coding and agentic tasks, the options are kimi-k2.6 (with vision support) and glm-5.1. For local use without an Ollama Cloud subscription, nemotron-3-super, gemma4:31b, and qwen3.6 are available. The vision support on kimi-k2.6 is relevant here: the browser annotation workflow depends on it.
If something goes wrong with your Codex App configuration, you can restore the previous state:
ollama launch codex-app --restore
Beyond the Codex App, the release includes one other notable change: the MLX sampler has been reworked for improved generation quality on Apple Silicon. No numbers are given, but if you run models locally on a Mac, generation behavior should feel different after upgrading.
What to do today: If you are running local AI workflows on a Mac or building with Ollama, upgrade to 0.24 and run ollama launch codex-app. Test the browser annotation flow against a local dev server. If you need cloud-grade reasoning, check kimi-k2.6 for tasks that benefit from vision input. If you are staying fully local, qwen3.6, nemotron-3-super, and gemma4:31b are your starting points.