LiteLLM Proxy for Claude Code

Problem

Claude Code only speaks the Anthropic API. If you want to use alternative LLM providers (GLM, Kimi, OpenRouter, etc.) or rotate multiple Claude OAuth tokens across a team, there is no built-in way to do it.

Solution

A LiteLLM proxy with a custom request transformer that converts Anthropic API calls into requests compatible with various providers. Claude Code connects to the proxy as if it were talking to Anthropic, and the proxy handles the rest.

Quick Start

1. Configure environment:

   cp .env.example .env
   cp config.yaml.example config.yaml

   Edit .env and fill in your API keys:

   ZAI_API_KEY=...
   KIMI_API_KEY=...
   OPENROUTER_API_KEY=...
   CLAUDE_OAUTH_TOKENS=sk-ant-oat-token1,sk-ant-oat-token2
   PROXY_STATIC_TOKEN=...
   LITELLM_MASTER_KEY=...

   Generate tokens: openssl rand -hex 32

2. Customize config.yaml for your deployment (add/remove models, adjust fallbacks).

3. Start the proxy:

   docker compose up -d

Claude Code Client Setup

Add to your Claude Code settings (~/.claude/settings.json or per-project .claude/settings.json):

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://your-server:4000",
    "ANTHROPIC_MODEL": "glm",
    "ANTHROPIC_CUSTOM_HEADERS": "x-litellm-api-key: YOUR_PROXY_STATIC_TOKEN",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

• ANTHROPIC_BASE_URL — point to your proxy instance
• ANTHROPIC_MODEL — the model alias from your config.yaml (e.g., glm, kimi, openrouter)
• ANTHROPIC_CUSTOM_HEADERS — passes the static auth token; the format is x-litellm-api-key: YOUR_PROXY_STATIC_TOKEN (no "Bearer" prefix)
• API_TIMEOUT_MS — request timeout in milliseconds (useful for slow models)
• CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC — disables telemetry and non-essential network requests

To skip the Claude Code onboarding screen, add to ~/.claude.json:

{
  "hasCompletedOnboarding": true
}

How It Works

Request Transformer

The proxy includes request_transformer.py — a custom LiteLLM callback that solves several compatibility problems when routing Anthropic API traffic to different providers:

1. Claude OAuth Token Rotation — Maintains a pool of Claude OAuth tokens (CLAUDE_OAUTH_TOKENS env var) and round-robins across them. Tokens that hit rate limits are automatically placed on a 60-second cooldown. When all tokens are cooling down, the one expiring soonest is used.

2. OAuth Header Stripping — Claude Code sends OAuth Authorization headers with every request. Non-Anthropic providers (GLM, Kimi) reject these headers. The transformer recursively strips them before forwarding.

3. Thinking Block Removal (Requests) — Some providers return "thinking" blocks in responses. Claude Code caches these in conversation history. When the conversation later routes to Anthropic, these cached blocks cause validation errors because Anthropic requires non-empty thinking content. The transformer strips thinking blocks from outgoing messages.

4. Thinking Block Removal (Responses) — OpenRouter returns reasoning/thinking blocks that would pollute the conversation cache. The transformer strips them from responses at the source.

Static Token Authentication

custom_auth.py provides lightweight authentication without a database. Set the PROXY_STATIC_TOKEN env var and all requests must include a matching token via the x-litellm-api-key header. If PROXY_STATIC_TOKEN is not set, all requests are allowed.

This is the "light mode" — no database required. For per-user tokens and usage tracking, you can optionally configure LiteLLM's built-in database-backed auth instead.

Claude OAuth Token Rotation

If your team uses multiple Claude Code Link sessions, you can pool the OAuth tokens for load balancing:

1. Each team member completes the Claude Code OAuth flow (Claude Code Link) to obtain an sk-ant-oat-... token
2. Collect the tokens and set them as a comma-separated list in the CLAUDE_OAUTH_TOKENS env var
3. The proxy rotates through tokens automatically, cooling down rate-limited ones

This allows the team to share Claude API capacity across multiple live sessions.

Token Mode Control

By default, the proxy rotates tokens from the CLAUDE_OAUTH_TOKENS pool for all Claude requests. You can switch to passthrough mode where each client's own session token is forwarded to Claude API as-is.

Global setting (env var):

CLAUDE_TOKEN_MODE=passthrough   # or "rotate" (default)

Per-request override (header): clients can override the global mode by adding x-claude-token-mode to their custom headers:

{
  "env": {
    "ANTHROPIC_CUSTOM_HEADERS": "x-litellm-api-key: YOUR_TOKEN, x-claude-token-mode: passthrough"
  }
}

Mode	Behavior
rotate	Server picks next token from CLAUDE_OAUTH_TOKENS pool (default)
passthrough	Client's own Authorization token is forwarded to Claude API

If mode is rotate but no tokens are configured in CLAUDE_OAUTH_TOKENS, the proxy automatically falls back to passthrough.

Provider Switching

The proxy routes requests to different providers based on the model name. Supported models are configured in config.yaml:

Model	Provider	Description
glm	ZAI API	GLM-5
glm-fast	ZAI API	GLM-5-AIR (faster)
kimi	Kimi API	Kimi-for-coding
openrouter	OpenRouter	Qwen 3.6 Plus (free)
anthropic/*	Anthropic	All Claude models via OAuth wildcard

Fallbacks: anthropic/* -> glm, anthropic/claude-haiku-* -> glm-fast

Configuration

• config.yaml — Model routing, fallbacks, and LiteLLM settings. Customized per deployment; see config.yaml.example for the reference configuration.
• .env — API keys and tokens. See .env.example for all required variables.
• request_transformer.py — Request/response transformation callback.
• custom_auth.py — Static token authentication handler.