MCP server for Garmin Connect — access activities, metrics, and FIT files via Claude Code
The server provides 18 well-named tools with clear action verbs (get-, list-, check-, download-) and good coverage of Garmin Connect's API surface. Most tools have descriptions (15/18 = 83%), but many are terse (10-50 chars) and lack actionable detail per the rubric's 50-200 char LLM-optimized baseline. Output schemas are NOT documented, descriptions don't explain what fields to expect, violating pattern:tool and pattern:response-shaper. Error handling exists (errorResult helpers) but is generic; no guidance on recovery or retryability. The tool garmin-login provides exceptional instructions (2000+ chars) but then 17 other tools are sparse by comparison. Naming is strong (verb_noun convention, clear intent), but composition has a gap: tools like get-daily-summary, get-heart-rate-variability, get-sleep-data, get-stress-data all accept the same 'date' parameter with identical defaults, these could benefit from a batch 'get-date-metrics' tool to reduce sequential calls. Security is good: no credentials in params (session stored server-side in ~/.garmin-connect-mcp/), but audit logging is absent.
Check if the saved Garmin Connect session is still valid. MUST be called after garmin-login to verify authentication worked.
Download the original FIT file for an activity. Returns the file path.
Returns step-by-step instructions for authenticating with Garmin Connect. Requires the Playwright MCP server to be installed. After following these steps, ALWAYS call the check-session tool to verify the login worked.
Get full activity summary (name, type, distance, duration, HR, calories, etc.)
Get time-series metrics for an activity (HR, cadence, elevation, pace over time)
Get heart rate time-in-zone breakdown for an activity
Output schemas not documented. Tools return jsonResult(data) but descriptions never explain what fields are in 'data', what structure to expect, or how downstream tools can chain. This violates pattern:tool (document output schema) and forces LLMs to guess at response structure.
Empty input schemas on 5 tools (garmin-login, check-session, get-user-profile have {} input). Code shows these are intentionally parameter-less (correct design), but the explicit empty schema {} should still be scored as 0 per the rule 'If a tool has NO input schema at all: its schema score MUST be 0.' These tools should still document their output.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Get full-resolution GPS track/polyline for an activity
Get lap/split data for an activity
Get weather conditions during an activity
Get blood pressure readings for a date
Get body composition (weight, BMI, fat %) for a date
Get daily summary: steps, calories, distance, intensity minutes, floors, etc.
Get heart rate variability (HRV) metrics for a date
Get sleep data for a date
Get daily stress and rest time data for a date
Get training status for a date
Get your Garmin Connect user profile and settings
List your Garmin Connect activities with pagination
Sparse descriptions on daily-metric tools. 15 tools have descriptions under 60 chars (e.g., 'Get daily summary: steps, calories, distance, intensity minutes, floors, etc.' is 77 chars but lacks LLM guidance on when to call it vs sibling tools). Rubric baseline is 50-200 chars for LLM-optimized descriptions; these hit the floor without context on differentiation, return types, or use cases.
No error recovery guidance. Tools return generic error messages (e.g., 'No Garmin session found. The user needs to run: npx garmin-connect-mcp login'). Per pattern:recovery-guide, errors should tell the LLM what to do next and classify retryability (is it transient? user-fixable? fatal?). No distinction between session-expired (retry login) vs API degradation (exponential backoff) vs rate-limit (wait and retry).
Repeated parameter definitions across 7 daily-metric tools. get-daily-summary, get-heart-rate-variability, get-sleep-data, get-stress-data, get-blood-pressure, get-body-composition, get-training-status all accept identical 'date' parameter with identical default. Violates composition pattern: agents will call these sequentially, wasting tokens and latency. No batch variant (e.g., get-daily-metrics taking metric_types[] and date).
No input validation documented. Parameters like 'limit' (1-100), 'start' (pagination offset), 'date' (YYYY-MM-DD), 'maxChartSize' (0+) have numeric ranges or format constraints, but descriptions don't explicitly state min/max or regex. Code likely validates internally, but LLM does not see constraints and may pass invalid values.
No audit logging or permission gates. Server doesn't log who called what tool, when, or what happened. Garmin Connect contains sensitive health data (HR, sleep, blood pressure, body composition). Per pattern:audit-trail, agent-initiated actions must be traceable for compliance. Also no permission checks (e.g., 'only the user whose session this is can call these tools').
Tool annotations missing. No readOnlyHint, destructiveHint, or idempotentHint decorations. Most tools are read-only (safe to retry), but download-fit is WRITE (downloads a file to disk). The LLM doesn't know which tools are reversible. Per 2026-07-28 spec, tool annotations help agents reason about side effects.