I found a model labeled azure/muse-special while Muse was building my website. So I dug deeper. This is Part 2 of digging through the Muse filesystem after my article hit the front page of Hacker News this week. In this article I focus on a model I found in my logs called muse-special, and confront the question: does Muse actually use Open AI and Claude models behind the scenes?

Nearly every session log in my VM was routed to Meta’s internal model, called Avocado. But one subagent used a model named azure/muse-special. The signature is tagged gpt_responses_v1 and contains an encrypted payload starting with g AAAAA (which Open AI uses). Tool call IDs used call_ followed by 24 mixed-case characters, different from the Avocado sessions.

The broader model catalogue shipped with Muse’s agent daemon lists about 15 versions of Avocado, plus: Claude Opus 4.6 / 4.7 / 4.8, Sonnet 4.6 and Haiku 4.5, GPT-5.5 and GPT-5.6 variants via Open AI, Azure and Codex, and Kimi K3 through Fireworks and Meta-hosted routes. The Claude support includes an Anthropic client with request handling, prompt conversion and streaming parsers.

There are API key files present for Anthropic, Open AI, etc., with access restricted to the inference-proxy service. But there’s also a proxy kill-switch setting in the env. The runtime has clients for multiple providers, which gives Meta the ability to change routing without asking users. In my case there was only a single outlier session that didn’t use the Avocado model, but the infra is there to do more.