Connect Claude Code through LiteLLM

Claude Code speaks only the Anthropic Messages API. The Fortytwo API serves Chat Completions, so Claude Code connects through a local LiteLLM proxy that translates between the two. This setup needs Python, Node.js and a terminal. Anthropic does not support Claude Code with models other than Claude.

Before you start: create a Fortytwo API key and buy credits — see Quickstart, step 1. Give each agent its own key with a Spend limit.

Setting	Value
Base URL	`https://api.fortytwo.network/v1`
Model ID	`qwen/qwen3.8-27b` in this guide, or another ID from Models

Install LiteLLM and Claude Code

pip install "litellm[proxy]==1.106.6"
npm install -g @anthropic-ai/claude-code

The LiteLLM version is pinned. Upgrade it on purpose: a new release can change how LiteLLM translates Messages requests.

Configure LiteLLM

Create litellm.yaml:

litellm.yaml

model_list:
  - model_name: qwen/qwen3.8-27b
    litellm_params:
      model: custom_openai/qwen/qwen3.8-27b
      api_base: https://api.fortytwo.network/v1
      api_key: os.environ/FORTYTWO_API_KEY

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

Use the custom_openai/ prefix: with openai/, LiteLLM sends Messages requests to the Responses API, which the Fortytwo API does not serve. The same file works for the Codex guide.The master key is a password for your local LiteLLM proxy; LiteLLM does not start without it. You do not get it anywhere: make up any string that starts with sk-, or let the command in the next step create a random one. Claude Code sends it to LiteLLM; LiteLLM sends your Fortytwo key to Fortytwo.

Start LiteLLM

export FORTYTWO_API_KEY="<your Fortytwo key>"
export LITELLM_MASTER_KEY="sk-$(openssl rand -hex 24)"
echo $LITELLM_MASTER_KEY
litellm --config litellm.yaml --port 4000

The echo line prints the master key: copy it for the next step. Keep this terminal open while you use Claude Code.

Start Claude Code

In a second terminal, in your project folder:

export ANTHROPIC_BASE_URL="http://localhost:4000"
export ANTHROPIC_AUTH_TOKEN="<the master key from step 3>"
export ANTHROPIC_MODEL="qwen/qwen3.8-27b"
export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen/qwen3.8-27b"
export ANTHROPIC_DEFAULT_SONNET_MODEL="qwen/qwen3.8-27b"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="qwen/qwen3.8-27b"
export CLAUDE_CODE_SUBAGENT_MODEL="qwen/qwen3.8-27b"
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=262144
claude

ANTHROPIC_AUTH_TOKEN replaces the sign-in: Claude Code does not ask for an Anthropic account.
The ANTHROPIC_DEFAULT_* and CLAUDE_CODE_SUBAGENT_MODEL lines send Claude Code’s background tasks and subagents to the same model. Without them, Claude Code asks LiteLLM for Claude models that it does not have.
CLAUDE_CODE_MAX_CONTEXT_TOKENS sets the model’s context window. Without it, Claude Code does not recognise the model and assumes 200,000 tokens.

Another model works the same way: take its ID, context window and max output from Models.On the first start, Claude Code asks for a text style and whether you trust the folder.

LiteLLM passes the model’s reasoning to Claude Code as thinking. More options: Claude Code LLM gateways , LiteLLM: Claude Code with non-Anthropic models .

Something not working with your agent, or using one we have not listed? Tell us in the platform: Feedback for ideas and small bugs, Contact support if it blocks your work, or Report a security issue.

Agents

Apps

Terminal

Advanced

Frameworks & SDKs

Connect Claude Code through LiteLLM