Anthroxy — AI skill for Claude Code

anthroxy

[](https://github.com/zekaizer/anthroxy/actions/workflows/ci.yml) [](https://github.com/zekaizer/anthroxy/releases/latest)

One endpoint for Claude Code in front of several backends. Point Claude Code at the router once; every model you configure shows up in its `/model` picker, and switching between them takes effect on the next request without restarting the session.

Claude Code ──► anthroxy ──┬──► vLLM              (Qwen, …)
 one URL,      routes by   ├──► LM Studio         (local models)
 one token     `model`     └──► api.anthropic.com (rotating OAuth token)

A backend can speak the Anthropic Messages API or the OpenAI Chat Completions API — the router translates the latter in both directions, streaming, tool calls, images and reasoning included, so Claude Code uses it like any other model.

Each backend carries its own credential: none, a static key, an environment variable, or a command that is re-run as the token expires. Claude Code only ever holds one static router token.

Everything that happens is visible: a request id on every response and log line, a browser console showing requests in flight and recently finished with their errors and hints, per-model statistics on disk, and an optional on-disk record of the exact bytes each request and response carried.

Install

**Linux x86_64 (WSL2 included), no toolchain needed** — every release ships a fully static executable:

curl -fsSLO https://github.com/zekaizer/anthroxy/releases/latest/download/anthroxy-x86_64-unknown-linux-musl.tar.gz
tar -xzf anthroxy-x86_64-unknown-linux-musl.tar.gz
install -m 755 anthroxy-x86_64-unknown-linux-musl/anthroxy ~/.local/bin/
anthroxy --version        # prints the version and the commit it was built from

**From source** (Linux and macOS, Rust 1.85+ for edition 2024):

cargo insta