Padwan Proxy

Padwan Proxy serves the Anthropic Messages API (/v1/messages and /v1/messages/count_tokens) on top of any OpenAI-compatible backend, so Anthropic clients such as Claude Code can use it. It is built on padwan-ai's Anthropic↔OpenAI translation layer and served by granian (RSGI) via gravier.

Prerequisites: Python 3.13+ with uv, and an API key for an OpenAI-compatible backend that supports function calling.

Quickstart

1. Export the backend key:

   export PADWAN_API_KEY=<YOUR_BACKEND_KEY>
   

2. Start the proxy:

   uvx padwan-proxy --backend-url https://api.example.com/v1/ \
     -m my-model --small-model my-small-model
   

   Or with Docker:

   docker run --rm -p 4000:4000 -e PADWAN_API_KEY \
     ghcr.io/polarsen-io/padwan-proxy:latest \
     --backend-url https://api.example.com/v1/ -m my-model --small-model my-small-model
   

3. Point the client at it:

   ANTHROPIC_BASE_URL=http://127.0.0.1:4000 ANTHROPIC_AUTH_TOKEN=dummy claude
   

Claude Code now answers with my-model. To keep model names in one place, let the proxy write the client config instead — see Claude Code.

Limitations

• The proxy does not translate Anthropic thinking budgets; reasoning must be enabled on the backend. Models that hide their reasoning emit no thinking blocks.
• count_tokens is an estimate (request size / 4), not a backend tokenizer count.
• The backend context window cannot be discovered; pass it with --context-window.