If you already use ChatGPT Desktop on a Mac and run models through Ollama, there is a new reason to consider bringing the two together. Ollama’s v0.34.0-rc1 release candidate adds a macOS integration for using Ollama models inside ChatGPT Desktop.
The useful question is whether it can keep the desktop workflow you like while giving you a local model for suitable jobs. My recommendation: consider a small trial if you are comfortable with preview software. Wait if you need a dependable replacement for your current work setup or a confirmed offline environment.
What has actually shipped?
GitHub lists the release on 5 September 2026. Although its heading says v0.34.0, the tag is v0.34.0-rc1 and the page is marked Pre-release. At the 6 September check, v0.33.3 remained the stable Latest release.
Ollama says setup is available through its macOS app. The release also lists improvements to structured outputs on Apple Silicon and compatibility work for tool search, response compaction and images. Those are release notes, not a promise of faster results for every model.
The implementation description says the new integration uses the existing ChatGPT/Codex profile, keeping access to chats and configured plugins, MCP servers and skills. That makes continuity the interesting part: can a local model handle a useful job in your familiar workspace?
The app, the model and the tools are separate choices
Keep three questions separate when evaluating this:
- Interface: can you find your conversation, files and controls?
- Model: can the selected model understand the task and produce a useful answer?
- Tools: can the full setup carry out an allowed action and return its result correctly?
A neat chat response only answers part of that test. For example, asking for a plan to reorganise notes is different from granting access to files and asking for the changes. Start with the plan. Then use disposable copies for any file-operation trial and inspect the result yourself.
Likewise, access to an existing skill does not prove that every local model can follow it reliably. Treat the integration as a way to evaluate a model in a workflow, rather than evidence that the model has gained the abilities of your usual hosted assistant.
Check which setup instructions you are following
There is an important documentation distinction. Ollama’s current Codex App guide describes an integration available since v0.24.0, using ollama launch codex-app. It documents persistent model selection, a restore command and backups of app configuration.
The new macOS release describes using the existing ChatGPT/Codex profile. Do not assume that the older command, backup location or restore behaviour applies unchanged. Before altering a working setup, match the instructions to the version and integration you intend to use.
OpenAI’s advanced configuration documentation confirms local-provider support for Codex, including Ollama. That general support does not settle the new integration’s exact account requirements or setup screens.
What is confirmed, and what still needs checking?
| Question | Current position |
|---|---|
| Which platform? | The new setup is announced for macOS. Do not infer Windows or Linux support from the availability of the desktop app itself. |
| Which release? | v0.34.0-rc1, a release candidate. |
| Which account or plan? | The checked sources do not establish a complete access matrix for this integration. |
| Does every tool work? | The implementation describes access to configured tools. Reliability needs checking with the selected model and task. |
| Is everything offline? | Local model execution does not, by itself, establish the behaviour of the surrounding app and connected tools. |
For hardware, Ollama’s general Mac requirements specify macOS Sonoma 14 or newer, with Apple M-series CPU/GPU support and x86 CPU-only support. Those are Ollama requirements, not a guarantee that every Mac meets the separate desktop app’s requirements or has enough memory for your chosen model.
Who should try it, and who should wait?
Consider trying it if you already have a local model you like, want to reduce app switching and can set aside time to check the result. Choose one modest job: summarise non-sensitive notes, review a small sample of code, or draft text from a short brief.
Wait if this is your only work machine and you cannot tolerate configuration changes, if you need a verified restore procedure, or if your reason for switching depends on a specific subscription being unnecessary. The available evidence does not yet support those assumptions.
If model size is your first obstacle, our llmfit guide is a useful starting point for exploring hardware fit. Keep capacity estimates separate from the question of whether a model completes your actual job.
A useful first evaluation
- Record the starting point. Note app versions, current provider settings and the supported way back before changing anything.
- Use a small, disposable task. Supply sample text or a copy of a test folder. Exclude personal and client data.
- Check model selection. Record the exact model, its local or cloud status and the displayed selection after restarting the app.
- Inspect the work. Check omissions, unwanted edits and tool results. A fluent answer is not enough.
- Check continuity and recovery. Confirm the intended chats and settings remain available, then verify the supported return to your previous setup.
FAQ
Does this make ChatGPT fully offline?
No such whole-app guarantee is established by the announcement. Ollama says it does not receive your prompts when you run locally, and its FAQ distinguishes local execution from cloud features. That statement covers Ollama’s handling; it does not establish how ChatGPT Desktop, a plugin or a remote service handles data. Check each part before using sensitive material.
Can I cancel my AI subscription?
Do not make that decision from this release alone. Check access requirements, the tasks you need, model quality and any connected services you still use. A local model trial gives you evidence for that decision; an integration announcement does not.
Is this the same as connecting Codex to another API?
It is a different setup path. Our DeepSeek V4 Flash and Codex Windows guide covers an external API provider. Keep its credentials and configuration instructions separate from this macOS Ollama integration.
Sources reviewed 6 September 2026. This preview covers Ollama’s macOS release candidate. Release status and documentation may change.