Claude Code Third Party Providers

Claude Code — Third-Party LLM Provider Connection

Connect OpenAI-compatible LLM APIs (Alibaba Qwen, DashScope, any /v1/chat/completions endpoint) to Claude Code, which only speaks the Anthropic Messages protocol.

When to Use

• User wants to use a non-Anthropic model (Qwen, GLM, DeepSeek, etc.) inside Claude Code
• User has an OpenAI-compatible API key/endpoint and wants Claude Code as the coding agent
• Claude Code says "Not logged in" or "Invalid API key" when pointed at a non-Anthropic endpoint

Core Problem

Claude Code only speaks the Anthropic Messages API format (/v1/messages with content blocks, tool_use, etc.). Setting OPENAI_API_KEY/OPENAI_BASE_URL does NOT work — Claude Code ignores them and demands Anthropic auth. Setting ANTHROPIC_BASE_URL to an OpenAI-compatible endpoint fails because the request/response formats differ.

Solution: Local Translation Proxy

A lightweight Python proxy on localhost translates:

• Inbound: Anthropic Messages API (POST /v1/messages) from Claude Code
• Outbound: OpenAI Chat Completions API (POST /v1/chat/completions) to the provider

Setup Steps

1. Write the proxy — use templates/proxy.py from this skill (copy to ~/claude-qwen-proxy/proxy.py)
2. Start the proxy:

   export OPENAI_API_KEY="<provider-api-key>"
   export OPENAI_BASE_URL="<provider-base-url>"   # e.g. https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1
   python ~/claude-qwen-proxy/proxy.py
   # Listens on http://127.0.0.1:8082
   

3. Run Claude Code through the proxy:

   export ANTHROPIC_API_KEY="sk-proxy-placeholder"   # ANY non-empty string works
   export ANTHROPIC_BASE_URL="http://127.0.0.1:8082"
   claude --bare --model qwen3.8-max-preview -p "Say hello" --max-turns 1
   

4. Verify — should get a text response. If "Invalid API key" → ANTHROPIC_API_KEY not set. If "model not found" → check provider's /v1/models endpoint for exact model IDs.

Windows Convenience Scripts

Create start-proxy.bat (sets OPENAI_* env vars, runs proxy.py) and claude-qwen.bat (sets ANTHROPIC_* env vars, runs claude). See templates/ directory.

Pitfalls

1. ANTHROPIC_API_KEY must be set — even to a dummy value. Without it, Claude Code enters OAuth login flow and ignores ANTHROPIC_BASE_URL.
2. --bare flag is important — skips OAuth/keychain reads, plugin sync, and CLAUDE.md discovery. Without it, Claude Code may try to validate the key against Anthropic's servers.
3. litellm proxy is an alternative but often fails on Windows — pip install litellm hits metadata generation errors. The custom proxy (stdlib-only, no deps) is more reliable.
4. Model names must match the provider's /v1/models list exactly — query it with curl -H "Authorization: Bearer $KEY" $BASE_URL/models.
5. Streaming translation is imperfect — tool_use blocks in streaming mode require stateful reassembly. The template proxy handles basic text streaming; complex multi-tool agentic loops may hit edge cases.
6. Proxy must stay running — it's a foreground process. Use start-proxy.bat in a minimized window or run under a process manager.
7. Claude Code features that depend on Anthropic-specific APIs won't work — e.g., extended thinking blocks, prompt caching headers, /usage command. Basic coding (read/write/bash) works fine.

Verification Checklist

# 1. Provider endpoint reachable?
curl -s "$OPENAI_BASE_URL/models" -H "Authorization: Bearer $OPENAI_API_KEY" | head -5

# 2. Proxy running?
curl -s http://127.0.0.1:8082/   # → {"status": "ok", "proxy": "anthropic-to-openai"}

# 3. Claude Code connected?
ANTHROPIC_API_KEY=sk-placeholder ANTHROPIC_BASE_URL=http://127.0.0.1:8082 \
  claude --bare --model <model-id> -p "Say hello" --max-turns 1

Known Working Providers

Gap Note

The bundled claude-code skill (v2.2.1) does NOT cover third-party provider connection. Its auth section only documents Anthropic OAuth, ANTHROPIC_API_KEY for Anthropic, and Bedrock/Vertex/Foundry. If that skill is ever updated, this content should be merged into it.