✨ Feature Request: Support for Self-Hosted LLMs in Claude Code Harness

🚀 Feature Request

Enable Claude Code to integrate with self-hosted LLMs, allowing users to swap out Claude’s proprietary models for locally run or open-source models (e.g., via OpenAI-compatible APIs, vLLM, LM Studio, Ollama, etc.).

---

🧠 Rationale

Anthropic currently offers two distinct, best-in-class products:

1. Claude LLMs (Claude 3 family, etc.) — a top-tier engine
2. Claude Code — a developer-focused “harness” or coding interface for interacting with those models

While these two are tightly integrated today, Claude Code itself is a product — and a highly capable one — that could be decoupled and monetized independently. By opening up support for self-hosted models, Anthropic can expand Claude Code’s market fit and revenue streams without compromising its brand integrity or compute strategy.

---

💡 Why This Matters

Right now, Claude Code is arguably the best code harness on the market: it’s fast, intuitive, multi-modal, and deeply contextual. But its use is locked to Anthropic models. Meanwhile, a growing community of:

• Developers with local inference setups (Ollama, GGUF, LM Studio)
• Enterprises with on-prem security requirements
• AI researchers needing LLM flexibility
• Startups building on open-source foundation models (e.g. DeepSeek, Qwen, Mistral, etc.)

…are actively seeking exactly this type of UI/harness — but cannot adopt Claude Code due to its closed coupling with Claude LLMs.

There’s a major market gap here. No one has built a best-in-class harness that is:

• IDE-like in productivity
• Model-agnostic
• Capable of supporting agentic workflows, prompt engineering, memory, etc.

**Claude Code can be that product.**

---

📈 Business Opportunity

Supporting third-party models doesn’t cannibalize Anthropic’s business — it expands it:

✅ Monetize the Harness

Offer Claude Code as a standalone paid product:

• Claude Code Pro (subscription-based): harness-only pricing for those using their own LLMs
• Claude Code Teams: collaborative, multi-agent coding environments, agnostic of backend model

✅ Reduce Compute Burden

By enabling local LLMs:

• Reduce server-side GPU load
• Improve margins
• Serve both ends of the spectrum: cloud users and local users

✅ Expand Developer Mindshare

Claude Code becomes the VS Code of AI coding harnesses — ubiquitous, community-extended, and sticky.

---

🛠 Suggested Implementation

• Allow users to configure Claude Code with:

• base_url, api_key, and model_id (OpenAI-compatible API pattern)
• Optional adapter for vLLM, Ollama, or similar
• Add a dropdown to toggle between Claude and custom backends
• Optional: Provide a plugin SDK for custom agents or tool integrations

---

🧪 Precedents & Momentum

Other companies (e.g., LM Studio, Continue, TabbyML, and VS Code plugins) have begun offering local inference support. However, none of them match Claude Code’s UX polish, deep context handling, or agentic flexibility.

Anthropic can own this space by shipping first-class, plug-and-play support.

---

🙌 Closing Thoughts

Claude Code is already world-class. Opening it to self-hosted LLMs would elevate it to category-defining — like GitHub for LLMs. Developers, researchers, and enterprises are eager to adopt Claude Code if it just lets them bring their own engine.

Please consider this feature request not as a "nice to have" — but as a strategic unlock for a new class of users.