claude-usage-monitoringlisted

# Reading Claude subscription usage programmatically The real, server-side usage numbers (what `claude` shows in `/usage`) come from an **undocumented** OAuth-authenticated endpoint. There is no published API; everything here was recovered by `strings`-grepping the Claude Code CLI binary. Treat it as fragile. ## Endpoint ``` GET https://api.anthropic.com/api/oauth/usage Authorization: Bearer <access_token> anthropic-beta: oauth-2025-04-20 ``` Response: buckets `five_hour`, `seven_day`, `seven_day_opus`, `seven_day_sonnet` (any may be `null`), each `{ "utilization": <0-100 number>, "resets_at": <ISO-8601 string> }`. There is also a `limits` array and `spend`/`extra_usage` objects. - **`utilization` is percent USED.** Remaining = `100 - utilization`. - **`resets_at` is an ISO-8601 string with microseconds** (`2026-06-27T16:19:59.398499+00:00`), NOT epoch seconds. `ISO8601DateFormatter` only handles milliseconds — strip the fractional part before parsing. - **HTTP 429** when polling too often — honor `Retry-After`, cache the last reading, poll on the order of minutes, not seconds. - Parse **leniently** (key-by-key, ignore unknown keys, missing buckets → nil). The shape will change. ## Getting a token Two options: 1. **Your own OAuth login (recommended for standalone tools).** Authorization-code + PKCE: - `client_id` `9d1c250a-e61b-44d9-88ed-5944d1962f5e` - authorize `https://claude.ai/oauth/authorize`, token `https://platform.claude.com/v1/oauth/token` -