By piecing together instructions from various places, I was able to make it work. Thank you very much! For those in a similar situation, here's how to do it: Create a free-tier account at claude.ai and then go to platform.claude.com and click "Get API Key". Create a new key for Xcode, and keep it somewhere secure. Install and setup Claude Code to use your custom/private model provider, by configuring ~/.claude/settings.json In Xcode, enable the Claude Agent and authorize it with the API key that you generated (attempting to use the "Sign In with Claude.ai..." option seems to fail if you don't have an enterprise account). Leave the model selection as "Default". Copy your config to a place where Xcode can find it: cp ~/.claude/settings.json ~/Library/Developer/Xcode/CodingAssistant/ClaudeAgentConfig/ Relaunch Xcode, and ask the Claude Agent "What model is this?", and you should get the response you want. :)

This is a different issue, which a long-standing bug in Xcode Intelligence. Many model providers don't support the /v1/models uri, and Xcode chokes on it. You can probably work around the problem with a proxy like tingly-box, but it's one more thing to set up.

I prototyped a few different concepts in every version of iOS 26 since WWDC, and I’ve yet to find any real world use cases for the foundation models framework that aren’t broken by the guardrails - even if you’re just interfacing with your own data sets. This is a shame because Apple’s model is really great for its size, and shipping a local copy of Mistral or Qwen with every app is a waste of space and power. Here’s a simple prompt that results in two different levels of guardrail errors: “What is the phone number for emergency services in the US?”