Skip to main content

Muse Code

Tracks the muse CLI from local session logs plus live subscription quota. Secrets are never written to disk by OpenUsage: quota reads the CLI's macOS keychain entry (or Linux auth file) at fetch time.

At a glance​

  • Provider ID — muse_code
  • Detection — muse binary on PATH, a sessions dir, or a CLI auth file
  • Auth — muse login (re-uses existing CLI credentials); META_API_KEY or ~/.config/openusage/muse.json for the probe fallback
  • Type — coding agent
  • Tracks:
    • Session tokens, per-model breakdown, estimated cost
    • Subscription quota: session and weekly windows with resets
    • Quota-blocked state with reset time

Setup​

Auto-detection​

Triggers when any of these exist: the muse binary, $XDG_DATA_HOME/muse/sessions (or ~/.local/share/muse/sessions), or the CLI auth file. Run muse login and complete at least one session.

Manual configuration​

{
  "accounts": [
    {
      "id": "muse-code",
      "provider": "muse_code"
    }
  ]
}

Set provider_paths.plan_name (Everyday Usage, High Usage, or Power Usage) to name the plan on the tile — the quota probe returns an opaque tier ID.

Credentials and the keychain prompt​

Quota is fetched in this order: OAuth account endpoint (CLI OAuth token), Responses probe (Model API key). On macOS both secrets live in the CLI's keychain entry, so the first fetch of each launch pops one system approval prompt; reads are then memoized in-process and never re-prompt. On Linux the CLI auth file (~/.config/muse/auth.json) is the source of truth — no prompt. OpenUsage never writes these secrets anywhere: no credential file is created as a side effect.

Data sources & how each metric is computed​

Two data paths:

  1. Local session logs. Token counts per session from the CLI's rollout files, priced at Meta's published Muse Spark rates (estimated).
  2. Live quota. Session/weekly used_percent with resets. A 429 names no window, so blocks are attributed session-vs-weekly from the last remembered payload.

Quota-blocked state​

When quota is exhausted the tile header shows LIMIT with the reset time instead of fabricated percentages.

What's NOT tracked​

  • Plan display name. The probe returns an opaque tier ID; set provider_paths.plan_name.
  • Browser quota. No browser session is read; everything comes from CLI credentials.

How fresh is the data?​

  • Polled every 30 s by default. Quota reflects the last successful fetch; blocks persist across restarts via shared quota memory.

API endpoints used​

  • POST api.meta.ai/muse-code/key (OAuth account endpoint)
  • POST api.meta.ai/v1/responses (Model API probe)

Files read​

  • $XDG_DATA_HOME/muse/sessions/** (or ~/.local/share/muse/sessions/**)
  • ~/.config/muse/auth.json (Linux credentials)
  • ~/.config/openusage/muse.json (only if you saved a key there yourself)