Remote MCP connector serving Garmin training data to Claude and ChatGPT. Reads live fitness activities from a private GitHub data repository and exposes them via a Cloudflare Worker.
Slipstream demonstrates solid tool design with clear verb-noun naming (list_activities, search_activities, data_status) and consistent parameter schemas using Zod. All 6 tools have descriptions (avg ~90 chars) and typed input schemas. However, output schemas are undocumented, responses are JSON-stringified text blobs without declared field structure. Parameter descriptions are minimal (e.g., sport_type lacks guidance on valid values beyond examples). No error handling guidance, no pagination metadata (total count, next_cursor), and no tool annotations (readOnlyHint). The server is read-only by design, which simplifies composition but limits real-world utility.
Totals (distance, time, elevation, calories, avg HR). Optionally group_by sport/month/year.
Check the fitness data is connected; returns count, date range, sources.
List activities, newest first. Filter by sport/date/name; sort and limit.
Distinct activity types with counts.
Activity-level bests: longest distance, longest time, most elevation, fastest pace.
Free-text search over activity name, type, and source.
Output schemas undocumented. All tools return JSON-stringified text responses without declared field structure. LLMs cannot plan downstream operations or extract typed data reliably.
Parameter descriptions lack format/constraint guidance. 'sport_type' says 'e.g. Run, Ride, Swim, Yoga' but does not state these are the only valid values or how to discover them. LLMs may hallucinate invalid sport types.
No pagination metadata returned. list_activities and search_activities accept limit but do not return total count or next_cursor. Large result sets cannot be properly paginated.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
No tool annotations. All tools are read-only but lack readOnlyHint annotation. Agents cannot distinguish safe tools from destructive ones without external documentation.
No error recovery guidance. Tools return raw errors (e.g., 'Could not read fitness data (GitHub 404)') without suggesting next steps or alternatives.