LiteLLM Claude Code Proxy

A LiteLLM proxy that routes any model request to your Azure AI Foundry deployment. Useful for running Claude Code against an AI model hosted on Azure Foundry.

Prerequisites

• uv installed
• An Azure AI Foundry deployment with an API key

Setup

1. Create the .env file

cp .env.example .env

Edit .env and fill in your values:

AZURE_OPENAI_ENDPOINT=
AZURE_OPENAI_API_KEY=
AZURE_OPENAI_DEPLOYMENT=azure/responses/gpt-5.3-codex
AZURE_OPENAI_API_VERSION=preview

LITELLM_PORT=4000
LITELLM_MASTER_KEY=sk-foo-bar-baz
LITELLM_LOCAL_MODEL_COST_MAP=true

UI_USERNAME=litellm
UI_PASSWORD=litellm

POSTGRES_USER=litellm
POSTGRES_PASSWORD=litellm
POSTGRES_DB=litellm

Variable	Description
AZURE_OPENAI_ENDPOINT	Your Azure AI Foundry endpoint URL
AZURE_OPENAI_API_KEY	API key for the endpoint
AZURE_OPENAI_DEPLOYMENT	Model deployment name with provider prefix (e.g. azure/responses/gpt-5.3-codex)
AZURE_OPENAI_API_VERSION	Azure API version (preview for Codex models, or a dated version like 2025-03-01-preview)
LITELLM_PORT	Port for the proxy (default: 4000)
LITELLM_MASTER_KEY	Auth token for the proxy — used by clients to authenticate against the LiteLLM proxy
LITELLM_LOCAL_MODEL_COST_MAP	Set to true to skip fetching the remote cost map (avoids SSL errors behind corporate proxies)
UI_USERNAME	Username for the LiteLLM admin UI at /ui (default: litellm)
UI_PASSWORD	Password for the LiteLLM admin UI at /ui (default: litellm)
POSTGRES_USER	PostgreSQL username (default: litellm, used by Docker Compose)
POSTGRES_PASSWORD	PostgreSQL password (default: litellm, used by Docker Compose)
POSTGRES_DB	PostgreSQL database name (default: litellm, used by Docker Compose)

Security note: The default values for LITELLM_MASTER_KEY, UI_PASSWORD, POSTGRES_USER, and POSTGRES_PASSWORD are not secure. If you expose this proxy beyond localhost (e.g. on a shared network or in production), replace them with strong, unique values.

2. Start the proxy

./start_proxy.sh

The proxy starts on port 4000 by default. Override with LITELLM_PORT in your .env.

3. Use with Claude Code

In a separate terminal:

export ANTHROPIC_BASE_URL=http://localhost:4000
export ANTHROPIC_AUTH_TOKEN=<your-LITELLM_MASTER_KEY>
claude

Configuring Claude Code

Claude Code needs two environment variables to connect through the proxy. You can set them in different ways depending on your workflow.

See also the official LiteLLM guide: Use Claude Code with Non-Anthropic Models.

Option A: Environment variables (per session)

Set them in your shell before launching Claude Code:

export ANTHROPIC_BASE_URL=http://localhost:4000
export ANTHROPIC_AUTH_TOKEN=<your-LITELLM_MASTER_KEY>
claude

Option B: Claude Code settings (persistent)

Run claude config or edit ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:4000",
    "ANTHROPIC_AUTH_TOKEN": "<your-LITELLM_MASTER_KEY>"
  }
}

This persists across sessions so you don't need to export variables each time.

Option C: VS Code settings (for Copilot Chat with Claude Code)

Add to your VS Code settings.json:

{
  "claude-code.env": {
    "ANTHROPIC_BASE_URL": "http://localhost:4000",
    "ANTHROPIC_AUTH_TOKEN": "<your-LITELLM_MASTER_KEY>"
  }
}

Choosing a model

By default Claude Code picks a model name like claude-sonnet-4-20250514. Since the proxy uses a wildcard, any model name is accepted and routed to your Azure deployment. You can also specify a model explicitly:

claude --model gpt-5.3-codex

4. Verify with curl

curl -X POST http://localhost:4000/v1/messages \
  -H "Authorization: Bearer <your-LITELLM_MASTER_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-20250514",
    "max_tokens": 100,
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

How it works

The litellm_config.yaml uses a wildcard (model_name: "*") so any model name sent by the client is routed to your single Azure deployment. The start_proxy.sh script uses uv run with UV_NATIVE_TLS=true to use the system certificate store, avoiding SSL issues behind corporate proxies.

Running with Docker

The Dockerfile uses the official LiteLLM Docker image — no need to install uv or Python separately.

Build the image

docker build -t litellm-claude-code-proxy .

Run the container

docker run --env-file .env -p 4000:4000 litellm-claude-code-proxy

Or skip the build entirely and run the official image directly:

docker run \
  --env-file .env \
  -v $(pwd)/litellm_config.yaml:/app/config.yaml \
  -p 4000:4000 \
  docker.litellm.ai/berriai/litellm:main-stable \
  --config /app/config.yaml --port 4000

The container reads all configuration from environment variables via the .env file.

Running with Docker Compose (Proxy + Database)

Docker Compose sets up both the LiteLLM proxy and a PostgreSQL database, which enables virtual keys, spend tracking, and other DB-backed features.

1. Configure .env

Make sure your .env includes the database variables (see .env.example):

POSTGRES_USER=litellm
POSTGRES_PASSWORD=litellm
POSTGRES_DB=litellm

The DATABASE_URL is assembled automatically from these variables in docker-compose.yml — you don't need to set it yourself.

2. Start the stack

docker compose up --build

This starts:

• litellm — the proxy on port 4000 (override with LITELLM_PORT)
• db — PostgreSQL 16 with a persistent pgdata volume

The proxy waits for Postgres to be healthy before starting.

3. Stop / clean up

docker compose down          # stop containers (data is preserved in the volume)
docker compose down -v       # stop containers and delete the database volume

References

• LiteLLM Documentation
• Proxy Quick Start
• Proxy CLI
• Proxy Config Settings
• Docker Quick Start & Model List Specification
• Use Claude Code with Non-Anthropic Models