Local, read-only Garmin Connect MCP server for Claude Desktop (stdio).
Garmin MCP demonstrates solid naming conventions (all tools start with action verbs: get_, list_, whoami) and comprehensive parameter schemas with proper type definitions and descriptions. All 18 tools have descriptions and input schemas are visible and well-structured using Zod. However, output schemas are not documented, the code returns JSON.stringify(data) without specifying what fields agents should expect. Parameter descriptions are adequate but generic (e.g., 'Date in YYYY-MM-DD format' repeated 18 times). Error handling is present but minimal, friendly() wraps errors with emoji prefixes but provides no recovery guidance or categorization. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all tools being read-only. Composition is strong: tools are single-responsibility and chainable (e.g., list_activities → get_activity). Security is good: no secrets in parameters, read-only operations only.
Detailed data for a single activity (GPS, heart rate, power, splits, weather, etc.). Pick via `include`: summary, splits, weather, all.
Body Battery timeline for a date range (energy level throughout the day).
Daily wellness metrics for a date. Pick one or more via `metrics`: summary, sleep, stress, heart_rate, hrv, spo2, respiration, hydration, steps, floors, intensity_minutes, body_battery, body_battery_events, stats_and_body.
Fitness/performance metrics. Pick via `metrics`: race_predictions, cycling_ftp, lactate_threshold, personal_records (all latest, date-agnostic), and endurance_score, hill_score, resting_heart_rate, weekly_intensity_minutes (over the given date range).
Floors climbed for a date range.
Heart-rate values for a date range (incl. resting heart rate).
Output schemas not documented. Tools return JSON.stringify(data) without specifying expected fields, types, or structure. LLMs cannot plan downstream operations or extract specific data reliably.
Error handling lacks recovery guidance. friendly() returns emoji-prefixed messages (🔒, ⏳) but does not categorize errors as retryable/user-fixable/fatal or suggest next steps. LLMs cannot self-correct.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
HRV (heart-rate variability) for a date range.
Hydration (fluid intake) for a date range.
Intensity minutes for a date range.
Respiration (breaths per minute) for a date range.
Sleep data for a date range (stages, duration, score).
Pulse Ox (SpO2) readings for a date range.
Step timeline for a date range.
Stress values for a date range.
Training metrics for a date. Pick via `metrics`: readiness, morning_readiness, status, vo2max, fitness_age.
Body weight & body composition (weight, BMI, body-fat %) over a date range. Set include_raw for the individual weigh-in entries too. Default range: the last ~4 weeks.
List the most recently recorded activities (paginated).
Check the Garmin connection and return the logged-in account profile (display name, full name).
No tool annotations despite all tools being read-only. Missing readOnlyHint in tool definitions prevents agents from understanding safety properties and optimizing execution strategy.
Generic parameter descriptions repeated across 13 tools. 'Date in YYYY-MM-DD format. Default: today.' and 'Start date (YYYY-MM-DD). Default: ~4 weeks ago.' appear verbatim, wasting tokens and reducing clarity. Should include context-specific guidance (e.g., 'Date of health metrics to retrieve').
list_activities lacks pagination documentation. Tool accepts limit and start parameters but does not document total count, next_cursor, or whether results are sorted. Agents cannot reliably paginate large result sets.