Differences between cloud and local models in Xcode Intelligence

I was surprised not to see a local model option after installing the Xcode 27 beta. I haven't watched any of the coding assistant videos yet, (very much not interested in cloud-based AI providers) but I am also curious about whether MLX has been found to be just not as good as the cloud-based providers, or whether there is some other reason Apple decided not to include a local-model option.

I'm excited to hear you're exploring this! We discussed this topic quite a bit in yesterday's Coding Intelligence group lab: https://www.youtube.com/live/Vd5nu5qIhGk?si=SG0yRbHTfHwy3VxA&t=2179

I'll also try to outline what is most useful below.

When you use the mode in Xcode that lets you provide just a quick way of accessing models, that will give you the ability to "chat" with the models, and those models will get a few tools, but to get the most done, you really want the much more advanced agentic features we shipped over the last few months. This will take the number of tools that your LLM can use from a few to about 50-60 inside Xcode, as well as asking your permission to run shell commands. At that point, the possibilities are pretty unlimited.

In order to do so much more, these agentic features need a much more sophisticated coordination system at their core. They rely on another system that runs on your computer and acts as a coordinator between the LLM and the tools it can use, managing its context window and keeping it on track. These are often called "agent harnesses" and when you click the "Get" button in Xcode next to Claude Agent, Codex, or Gemini, this is what you're installing.

When commercial providers like Anthropic, OpenAI, and Google build agentic tools, it's normal for them to provide their own harness. But you are in a special situation with on-device models. You may want to connect to a different harness.

In the WWDC session on local agentic AI (https://developer.apple.com/videos/play/wwdc2026/232/), we show how to get set up with OpenCode, a popular open coding agent harness, on the command line. This is a great option to get started with, because once you are up and running, it is also immediately portable to Xcode 26.6 and Xcode 27 with ACP.

To do so:

1. Click "Add an agent..." in the Xcode's intelligence settings.
2. Go back to your command line and run which opencode.
3. Paste the path that you see into the "Executable" blank in the configuration.
4. Add acp (no dashes or slashes) as the only argument in the "Arguments" list in settings.

Now, when you make a new conversation and choose your OpenCode agent, it will have access to all of Xcode's tools and give you full agentic functionality.

Hope this helps!

what features am I missing out on by using a local model instead of a larger, cloud-based model?

In terms of what Xcode exposes, I think that there is feature parity. As long as your LLM can communicate via the supported API (for chat mode), or the agent harness speaks ACP (for agentic mode), Xcode doesn't care where the model is running and, afaik, it doesn't enable any extra controls if you're using a cloud model compared to what is available when you use a local model.

Big caveat, though: Xcode uses prompts that have likely been tested mostly with Claude and GPT (and now Gemini), and may not work well at all with your local model. There is no way currently to replace them (Apple engineers please correct me if I'm wrong). I wish there was a possibility for “advanced” users to tweak that.

Is Xcode's ability to run tests, build the project, etc. limited by the use of a local model? Does it limit the ability to do more complex agentic workflows? If so, how?

In Xcode, if you enable Settings > Intelligence > Allow external agents to use Xcode tools, an agent supporting MCP can control Xcode directly, and in my experiments a good local model will build and test Xcode projects just fine (and iterate in case of errors). Even if you don't enable MCP, the agent will probably try to use xcodebuild, xcrun, etc anyway, although the user experience is not as good because Xcode won't display the changes as it would otherwise do.

Regarding the quality and complexity of agentic workflows, the key point is not Xcode, but the combination harness+model (+hardware, but that's not something one can usually change as often). Imo you should first get that working well outside of Xcode and then integrate it in Xcode, which may require additional tuning (and ACP compatibility—most ACP-enabled harnesses I've tried currently do not work with Xcode for some reason).

what cloud-based models can offer that I don't already have.

More computational power, hence more “intelligence” (scope, quality, performance). Whether that's necessary to you depends only on your needs. They may be also easier to set up in Xcode.

Running a local model might also be less cost-effective at the moment, but I'm not sure. I'd rather spend a few more $$$ of electricity in any case to be in control.

Nothing else. Environmentally, the best option is probably to avoid using anything. Ethically, many lines have been crossed already, but I believe that “open” models are the lesser evil going forward.