muse-sub-proxy (standalone)

Expose Meta API access through OpenAI-compatible endpoints. No muse CLI anywhere — pure Python stdlib talking direct REST to https://api.meta.ai/v1.

External tool (OpenCode/Hermes/etc)     muse-sub-proxy              Meta API
  POST /v1/chat/completions  --->  POST /v1/responses (Bearer LLM|) --->  Muse Spark
  OpenAI JSON (+SSE)         <---  output_text translated     <---

Setup

No CLI needed at runtime. You need one Meta API key (LLM|...), resolved as:

1. MUSE_SUB_API_KEY env var (Linux/portable), or
2. macOS keychain item ai.meta.dev.credentials / account meta (read via
   security CLI — one-time Always-Allow approval), or
3. ~/.config/muse-sub-proxy/api_key.txt (mode 0600).

Run:

python3 proxy.py                        # :8920
SUB_PROXY_PORT=8921 python3 proxy.py    # custom port
curl -s localhost:8920/v1/models

LaunchAgent: cp com.muse-sub-proxy.plist ~/Library/LaunchAgents/ && launchctl load ...

Endpoints

• GET /v1/models — live ids from Meta (statics are fallback only, so new
  models survive CLI/API updates with no proxy change)
• GET /v1/usage — latest subscription quota snapshot (see below), persisted
  to ~/.config/muse-sub-proxy/usage.json after every chat completion
• POST /v1/chat/completions — OpenAI chat format, stream: true/false.
  Both modes return the token usage object; both also carry an extra
  subscription block (quota snapshot for that exact call).

OpenCode / Hermes example

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "muse-sub": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Muse Sub",
      "options": { "baseURL": "http://127.0.0.1:8920/v1" },
      "models": {
        "muse-spark-1.2": { "name": "Muse Spark 1.2 (sub)", "limit": { "context": 1048576, "output": 131072 } }
      }
    }
  }
}

Hermes: provider entry muse-sub with base_url: http://127.0.0.1:8920/v1
(set via hermes config set providers.muse-sub.base_url ...).

Reverse-engineering notes (how we got here)

• Subscription page says "10–50 prompts every 5 hours" — weighted user
  prompts, each fanning out into dozens/hundreds of internal model calls.
• The muse launcher (~/.local/bin/muse, plain bash) contains the whole
  OAuth device flow in cleartext: https://auth.meta.com +
  `/oidc/device/authorizatio