Claude Code Wrapper

Anthropic Messages API wrapper that exposes Claude Agent SDK as a standard API endpoint.

Direct (Anthropic API):
┌────────┐  /v1/messages  ┌──────────────┐       ┌───────────┐
│ Client  │──────────────▶│ Wrapper:8790 │──────▶│ Claude SDK│
└────────┘                └──────────────┘       └───────────┘

With LiteLLM (OpenAI API compatible):
┌────────┐  /v1/chat/completions  ┌─────────────┐  /v1/messages  ┌──────────────┐       ┌───────────┐
│ Client  │──────────────────────▶│ LiteLLM:4000│──────────────▶│ Wrapper:8790 │──────▶│ Claude SDK│
└────────┘                        └─────────────┘               └──────────────┘       └───────────┘

Features

• Anthropic Messages API (/v1/messages)
• Model listing (/v1/models)
• Adaptive thinking with effort=high (internal, not configurable via API)
• External tool proxy (client tools → MCP stdio proxy)
• MCP server configuration
• Hash-based automatic session matching
• SSE streaming
• Multi-arch support (amd64 / arm64)

Quick Start

1. Generate OAuth Token

On a machine where you're logged into Claude:

claude setup-token

Copy the token (valid for ~1 year).

2. Create docker-compose.yml

services:
  claude-code-wrapper:
    container_name: claude-code-wrapper
    image: ghcr.io/farewellsagittarius/claude-code-wrapper:latest
    ports:
      - "8790:8790"
    environment:
      - INTERNAL_API_KEY=sk-claude-code-wrapper
      - CLAUDE_CODE_OAUTH_TOKEN=${CLAUDE_CODE_OAUTH_TOKEN}
      - TOOLS=Read,Grep,Glob,WebSearch,WebFetch
      - LOAD_USER_MCP=false
      - LOG_TO_FILE=false
    restart: unless-stopped

3. Create .env

CLAUDE_CODE_OAUTH_TOKEN=sk-ant-oat01-xxxxx

4. Start

docker compose up -d

Local Development

cp .env.example .env
# Edit .env and set CLAUDE_CODE_OAUTH_TOKEN
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -m src.main

Configuration

Variable	Description	Default
INTERNAL_API_KEY	Wrapper API auth key	sk-claude-code-wrapper
CLAUDE_CODE_OAUTH_TOKEN	Claude OAuth token (via claude setup-token)	(required)
PORT	Server port	8790
HOST	Listen address	0.0.0.0
TOOLS	Tool config: unset=all, ""=none, "Task,Bash,Read"=specific	all
LOAD_USER_MCP	Load user MCP servers from ~/.claude.json	true
EXPOSE_THINKING	Pass <thinking> blocks through to client	false
CLAUDE_CWD	Claude working directory	temp dir
DEBUG_MODE	Debug mode	false

Usage

from anthropic import Anthropic

client = Anthropic(
    base_url="http://localhost:8790/v1",
    api_key="sk-claude-code-wrapper"
)

response = client.messages.create(
    model="claude-code-opus",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.content[0].text)

Model Aliases

Name	SDK Model
claude-code	SDK default
claude-code-opus	opus
claude-code-sonnet	sonnet
claude-code-haiku	haiku

API Endpoints

Endpoint	Method	Description
/v1/messages	POST	Anthropic Messages API
/v1/models	GET	Model listing
/v1/sessions	GET	Active sessions
/health	GET	Health check
/docs	GET	Swagger docs

License

MIT