muse.ai, reverse-engineered into an OpenAI-shaped API

What it does muse2api wraps the muse.ai web app in an OpenAI-compatible REST API. It speaks /v1/chat/completions (plus the newer /v1/responses), /v1/images/generations and /v1/images/edits, and an async video task API for text-to-video and first-frame image-to-video. Under the hood it’s a headless-browser automation rig: CDP protocol, warm WebSocket tunnels, and session management keep the upstream web session alive and responsive.

The interesting bit The engineering is mostly in the plumbing — a multi-account pool with warm-tab affinity and LRU scheduling, per-account isolation via Chromium’s Target.createBrowserContext incognito containers, automatic 48-hour session renewal, and a conflict scheduler so the same account never steps on itself. A companion Chrome extension (or Tampermonkey script) exports login cookies, including HttpOnly ones, straight into the account pool.

Key highlights

• Claims 2–3 second first-token latency on streaming chat via hot-tab and WebSocket tunnel reuse.
• Model alias mapping routes familiar names like gpt-4o, gpt-5, claude-sonnet-4, deepseek-chat to Muse’s models.
• Video generation supports 5/6/8/10-second durations (actual length depends on upstream) and 9:16 or 16:9 framing, with a built-in media endpoint for generated MP4/WebP files.
• Per-account HTTP/SOCKS5 proxies; proxy failures raise errors rather than silently falling back to direct connection.
• Dark-mode admin console with account pool status, task replay, media library, and a one-click online upgrade that preserves local config.

Caveats

• This is reverse-engineering a commercial service — expect it to break whenever muse.ai changes its frontend, and mind the terms-of-service implications of account pooling.
• Default concurrency is 2; scaling up depends on your machine running multiple Chromium contexts.
• The README notes session renewal only stores what upstream actually returns — revoked or restricted accounts still need manual re-import.

Verdict Worth a look if you want a self-hosted OpenAI-compatible gateway over muse.ai accounts and accept the fragility that comes with scraping a moving target. Skip it if you need a stable, sanctioned API — this is glue code with a shelf life.