Does this only support Claude Code? Supporting multiple LLM providers would be nice!

[YehyunLee]

As the title suggests, it seems like the app currently only supports Claude Code.
It would be nice if this could support multiple LLM providers.

If that's the case, I would be happily move from open code to this!

[jordan-BAIC]

@YehyunLee The plan is to eventually support multiple agent providers, the reason I haven't yet is that it seems that Claude Code is a clear best-of-class in both capabilities and pricing.

I haven't really spent much time using other providers though, so if you make a compelling case that this is a worthwhile feature right now I would move it up the priority list.

[YehyunLee]

Thanks ! Yeah I only asked because Claude Code rate limit is somewhat bad. Perhaps supporting Google will be superb!

BTW, this is only for Mac and Linux right? No support for Windows yet

[mattiapintus]

@jordan-BAIC @YehyunLee I just created a Feature Request post related to this matter: #206

I'd like to add some compelling points for why local LLM support specifically (LM Studio, Ollama, etc.) would be valuable right now:

The compelling case:

1. Rate limits (as @YehyunLee mentioned) - Local models have zero rate limits. With Crystal's multi-session approach, users could easily hit API limits during intensive development sessions.

2. Cost for experimentation - Crystal's strength is running multiple parallel sessions. This is amazing but can get expensive fast when testing different approaches. Local models = unlimited experimentation at zero cost.

3. Privacy-first development - Many companies prohibit sending proprietary code to external APIs. Local LLM support would make Crystal enterprise-friendly.

4. Offline capability - Developers working on planes, trains, or with unstable internet could continue using Crystal.

5. Technical feasibility - Most local LLM servers (LM Studio, Ollama) already provide OpenAI-compatible APIs, so implementation could leverage existing code with minimal changes.

While Claude Code is indeed best-of-class for capabilities, local models serve a different use case - they're not competing with Claude Code, they're complementary. Users could use Claude Code for complex tasks and local models for rapid iteration/testing.

The full feature request has implementation details and mockups if you're interested. I'm open to discuss it further!