Ollama Can Now Launch the Codex App With Local Models
Ollama v0.24.0 can now launch and configure the installed Codex App against Ollama local or cloud models.
That is the news. The useful part is what it changes for operators.
Codex CLI already worked with Ollama. That path was for terminal-native users. The new target is the downloadable Codex App: the desktop coding agent OpenAI offers for macOS and Windows. OpenAI documents the app in the Codex quickstart. Ollama's v0.24.0 release includes the support, and the commit is explicit: "docs: add codex app".
The result is simple: the desktop app can use Ollama's OpenAI-compatible endpoint. That means local models when the workstation can handle them, and Ollama Cloud models when the local machine is not enough.
Ollama Cloud is still cloud inference. It is not private, offline, or local just because it routes through Ollama. The point is choice: one app surface, more model paths.
The command
Prerequisite: install and open the Codex App for macOS or Windows first. Ollama is configuring the installed desktop app, not installing Codex for you.
ollama launch codex-app
To launch with a specific model:
ollama launch codex-app --model gemma4:31b
For an Ollama Cloud model:
ollama launch codex-app --model kimi-k2.6:cloud
Those model names come from Ollama's docs. Availability still depends on what you have pulled locally and what your Ollama Cloud account can access.
To undo the change and restore the prior Codex App profile:
ollama launch codex-app --restore
Ollama says the launch is persistent until restored. The model stays selected the next time Codex opens. That is good operator behavior, but it is still state mutation. Know that before you run it.
Ollama also keeps the Codex CLI profile separate from the Codex App profile. That matters if you already have a terminal setup you do not want the app path to disturb.
Why this matters
First, it moves local models into a desktop coding-agent surface. A terminal agent is powerful, but it filters out users who do not want shells, config files, and command history as their main interface. The Codex App gives them repository browsing, app-based review, and browser annotation without giving up Ollama routing.
Second, it makes review less awkward. Coding agents are not done when they produce a diff. The loop is: inspect the diff, comment on the exact change, run another pass, verify again. If the desktop app makes that loop easier while Ollama owns the model endpoint, that is a better control surface for many repo-local tasks.
Third, it keeps the model decision separate from the interface decision. Use a local model for small edits, repo exploration, generated tests, copy changes, and offline work. Use an Ollama Cloud model when the workstation is weak but the Ollama path is convenient. Use a frontier hosted model when the task has high blast radius, long context, or migration risk.
The win is not that local models beat everything. They do not. The win is that Codex App no longer forces one model path.
The recovery path matters
Ollama's docs say it saves backups before overwriting Codex App config files. The backup path is ~/.ollama/backup/codex-app/. On Windows, ~ resolves to the user's profile directory.
That detail is worth calling out. App integrations fail when they mutate state without a clean undo path. --restore and a known backup directory make this safe enough to test on a real workstation.
The bigger pattern
The local-agent market is splitting by operator surface:
- CLIs for fast users who want direct file and shell control
- TUIs for long coding sessions where the terminal is the cockpit
- desktop apps for review, browsing, annotation, and lower-friction workflows
Ollama already had a strong place in the first two. Codex App support gives it a cleaner path into the third.
That is the real update: Ollama is becoming less of a local-model runtime you call from other tools and more of a routing layer for coding-agent surfaces.
If you are evaluating local agent workflows, start by deciding where control needs to sit: model, runtime, files, tools, logs, or human review. This update gives you one more way to split those decisions cleanly. If you want help designing that split for your team, start with Mimir Works services.
Read next: The Best GUIs, TUIs, and CLIs for Running AI Agents Locally and What It Actually Costs to Run AI Agents: A Practical Breakdown.