Ollama 0.24 ships support for the Codex App, OpenAI's desktop experience for running Codex threads in parallel. It includes built-in worktree support and git functionality. Launch it with one command:
ollama launch codex-app
That is the full setup. No extra configuration needed to get started.
What the app actually gives you
Three features stand out for builders.
First, a built-in browser. Codex can load local servers and sites directly inside the app. You can annotate on the page to request changes, which cuts the round-trip between browser, editor, and agent prompt.
Second, a review mode. You can read generated code, leave inline comments, and iterate without switching context. The whole loop stays inside one workspace.
Third, parallel thread support with worktrees. This matters when you have multiple tasks running at once and want git isolation between them.
Model choices
The release lists recommended models by use case. For difficult coding and agentic tasks, the source points to kimi-k2.6 (which includes vision support) and glm-5.1. For local use without an Ollama Cloud subscription, the options are nemotron-3-super, gemma4:31b, and qwen3.6.
Vision support in kimi-k2.6 is relevant here. The built-in browser annotation feature benefits directly from a model that can interpret what it sees on screen.
Restoring your previous setup
If you try the Codex App and want to roll back, the restore flag handles it:
ollama launch codex-app --restore
This brings back your previous configuration without manual cleanup.
Also in 0.24
The release includes a reworked MLX sampler for improved generation quality on Apple Silicon. If you run Ollama on a Mac and care about output consistency, this is worth noting even if the Codex App is the headline feature.
What to do with this today
If you are building or prototyping on a local machine, pull 0.24 and run ollama launch codex-app. Try the browser annotation flow against a local dev server. The combination of visual feedback, inline review, and worktree isolation covers most of the friction points in a typical frontend iteration loop. If you hit issues or want to compare outputs, the model list gives you clear options depending on whether you have cloud access or want to stay fully local.