Skip to content

ALCF-proxy

PyPI version Python versions License: MIT

alcf-proxy is a standalone, cross-platform translating gateway for ALCF inference endpoints. Run one local proxy and point any OpenAI- or Anthropic-style coding CLI at it — Codex, Claude Code, Aider, OpenCode, and more — without each tool needing to know about ALCF's auth or wire format.

Inspired by argo-proxy, but the upstream is ALCF's public Bearer-authenticated gateway (not ANL ARGO), and the proxy manages ALCF's rotating token for you.

Why

ALCF's gateway is OpenAI-compatible, but each coding CLI speaks a slightly different wire format and none of them know how to fetch/refresh an ALCF Globus token. alcf-proxy sits in the middle:

  your CLI  ──►  alcf-proxy (localhost)  ──►  ALCF inference gateway
  (chat / responses / messages)   translate + inject token, stream back
  • One instance, three client formats — /v1/chat/completions (OpenAI SDK, Aider, OpenCode), /v1/responses (Codex), /v1/messages (Claude Code).
  • Real streaming — every request is translated to ALCF's streaming chat/completions upstream and streamed back token-by-token in the client's format (via llm-rosetta).
  • Built-in auth — one-time Globus browser login; the proxy caches and auto-refreshes the ~48h access token. Clients just use a dummy key.
  • Endpoint failover — if a cluster is under maintenance, the proxy fails over to the next one and remaps the per-endpoint model id automatically.
  • Cross-platform — Linux, macOS, Windows. Pure Python.

60-second start

pip install alcf-proxy
alcf-proxy auth login     # one-time Globus browser login
alcf-proxy serve          # gateway on http://localhost:11445

Then point your CLI at http://localhost:11445/v1 with any dummy API key. See Quickstart and Connect a CLI.

Next steps