An MCP server that wraps the Oura Ring v2 API (sleep, readiness, activity, HRV, SpO2, stress, and more) using FastMCP.
16 tools with consistent naming (verb_noun pattern: get_*), clear descriptions (avg ~150 chars), and explicit input schemas. All parameters have type and description. However, output schemas are not formally documented in the code, responses are inferred from the _collection() helper. No error handling guidance, no tool annotations (readOnlyHint), and no pagination support despite tools returning potentially large datasets (e.g., get_heart_rate time series). Descriptions are good but could be more prescriptive about when to use each tool vs. similar ones.
Daily Activity scores: steps, active/total calories, MET minutes, activity levels, and the contributor breakdown for each day.
Daily Cardiovascular Age estimate (vascular age vs. chronological age).
Daily Readiness scores and contributors (activity balance, body temp, HRV balance, resting heart rate, recovery index, sleep balance).
Daily Resilience: long-term stress/recovery capacity level with sleep-recovery, daytime-recovery, and daytime-stress contributors.
Daily Sleep scores and contributors (deep sleep, efficiency, latency, REM, restfulness, timing, total sleep) for each night in the range.
Nightly average blood oxygen (SpO2) percentage and breathing disturbance index.
No output schema documentation. Tools return dict with 'data' array, 'count', 'endpoint', and date range, but LLMs cannot see the structure of individual data items (sleep stage durations, HRV values, etc.). This forces LLMs to guess field names and types.
No pagination support. Tools like get_heart_rate (time-series samples) and get_sleep_periods can return hundreds or thousands of items. No limit, offset, or cursor parameters. Large responses will exhaust context windows.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | <=2025-11-25 | v2 |
Daytime stress summary: seconds spent in high-stress vs. recovery, and an overall day summary label.
Time-series heart rate samples (bpm) with source (awake/sleep/rest/workout) and timestamp. Uses ISO 8601 datetimes; defaults to the last 24 hours. Windows can be large; this is raw samples, not a daily summary.
The user's profile: age, biological sex, height, weight, and email as recorded in Oura.
Rest Mode periods (when the user flagged illness/recovery) and their episodes.
Moment / meditation / rest / breathing sessions with HR, HRV, and motion-count time series.
Detailed sleep periods (the richest sleep resource): per-period stages in seconds, HR/HRV time series, respiratory rate, latency, efficiency, and the sleep hypnogram. There can be more than one period per night.
Recommended optimal bedtime windows and sleep-timing guidance.
Enhanced Tags: user-logged notes/events (e.g. caffeine, alcohol, naps, symptoms) with timestamps and optional comments.
VO2 max (cardiorespiratory fitness) estimates in the date range.
Recorded workouts: activity type, intensity, calories, distance, and start/end times (auto-detected or imported from Apple Health / Google Fit).
No error recovery guidance. Error responses return {'error': str(exc), 'endpoint': endpoint} with no actionable next steps. LLMs cannot distinguish retryable errors (rate limit, timeout) from user-fixable ones (invalid date format) from fatal ones (auth failure).
No tool annotations. All tools are read-only but lack readOnlyHint annotation. This prevents agents from understanding which tools are safe to call speculatively vs. which require confirmation.
Descriptions lack differentiation. Many tools (get_daily_sleep, get_daily_readiness, get_daily_activity, etc.) have similar descriptions that don't explain when to use one vs. another. LLMs may conflate them or pick the wrong tool.