anthroproxy

Protocol translation proxy: accepts Anthropic Messages API format and forwards to any OpenAI-compatible endpoint.

Use it to make tools that only speak Anthropic's API (like Claude Code) work with OpenAI-compatible backends (OpenAI, DeepSeek, Ollama, vLLM, etc.).

Quick start

npx anthroproxy \
  --target-url https://api.openai.com/v1 \
  --api-key sk-xxxx \
  --port 8080

Or with a .env file:

echo "TARGET_URL=https://api.openai.com/v1
TARGET_API_KEY=sk-xxxx" > .env
npx anthroproxy

Usage with Claude Code

claude config set proxy http://localhost:8080

Now Claude Code sends requests to the proxy, which translates and forwards them to your target backend.

Options

CLI flags

| Flag | Default | Description | |---|---|---| | --port <n> | 8080 | Port to listen on | | --target-url <url> | https://api.openai.com/v1 | Upstream API base URL | | --api-key <key> | — | API key for the upstream | | --model-mapping <str> | — | Model mapping, e.g. claude-3:gpt-4o,claude-4:gpt-4-turbo | | --default-model <name> | gpt-5 | Fallback model if no mapping matches | | --log-level <level> | DEBUG | One of: DEBUG, INFO, WARN, ERROR, NONE | | --help | — | Show help message | | --version | — | Show version number |

Environment variables

All CLI flags have matching environment variables:

| Variable | Corresponding flag | |---|---| | PORT | --port | | TARGET_URL | --target-url | | TARGET_API_KEY | --api-key | | MODEL_MAPPING | --model-mapping | | DEFAULT_MODEL | --default-model | | LOG_LEVEL | --log-level |

Model mapping

When the incoming request uses an Anthropic model name, the proxy maps it to a target model. Without a mapping, it falls back to --default-model.

--model-mapping "claude-sonnet-4-20250514:gpt-4o,claude-opus-4:gpt-4-turbo"

Endpoints

| Path | Description | |---|---| | POST /v1/messages | Translate and forward Anthropic → OpenAI request | | GET /health | Health check (returns {"status":"ok","uptime":...}) |

Features

• Streaming & non-streaming — Full SSE event translation
• Extended thinking — Native thinking content block support
• Tool use — Bidirectional tool_use ↔ function calling conversion
• Image input — Anthropic image blocks → OpenAI image_url
• Reasoning fallback — Automatic retry without reasoning_content if upstream doesn't support it

License

MIT