Bring your Garmin data to AI. An MCP server that bridges the garmin-grafana InfluxDB ecosystem with AI agents.
概览
An optional, self-hosted Model Context Protocol (MCP) server that exposes your Garmin training data to AI assistants. It is designed as a companion add-on to the arpanghosh8453/garmin-grafana project and reads directly from the InfluxDB instance that garmin-grafana already manages. It returns raw numbers from InfluxDB. All interpretation, analysis, and planning is delegated to the AI model consuming the tools — this server makes no assumptions about your fitness goals. Because this MCP server provides structured, context-aware data to LLMs like Claude, it acts as a world-class personal coach. Here is an actual, unedited response from Claude (via Copilot) summarizing a ride by cross-referencing simultaneously (Activity Details, Training Status, Sleep/Recovery, Training Zones, and Stress): "I just came back from my long Z2 ride. Please assess it in detail against our training plan!" View Claude's Analysis (Click to expand) - : HR locked at 121–122 bpm.
README
以下为原文 README 节选，完整内容请查看原文。
garmin-grafana-mcp-server
An optional, self-hosted Model Context Protocol (MCP) server that exposes your Garmin training data to AI assistants. It is designed as a companion add-on to the arpanghosh8453/garmin-grafana project and reads directly from the InfluxDB instance that garmin-grafana already manages.
This server is a pure data access layer. It returns raw numbers from InfluxDB. All interpretation, analysis, and planning is delegated to the AI model consuming the tools — this server makes no assumptions about your fitness goals.
Table of Contents
- How it fits into the ecosystem
- 🪄 See it in action: The AI Coach
- Prerequisites
- Quick Start (Docker Compose)
- Deployment (Docker — recommended)
- Local Development
- Connecting to an MCP Client
- MCP Tools
- Example prompts
- Configuration reference
- Project structure
- Schema compatibility
- Troubleshooting
- License
How it fits into the ecosystem
Garmin Device
│
▼
garmin-grafana ← you already have this running
├─ Garmin API sync
├─ InfluxDB (stores your data)
└─ Grafana (dashboards)
garmin-grafana-mcp-server ← this repo adds this
└─ Reads from the same InfluxDB
└─ Exposes MCP tools to AI clients (Claude, Copilot, etc.)
🪄 See it in action: The AI Coach
Because this MCP server provides structured, context-aware data to LLMs like Claude, it acts as a world-class personal coach.
Here is an actual, unedited response from Claude (via Copilot) summarizing a ride by cross-referencing six different MCP tools simultaneously (Activity Details, Training Status, Sleep/Recovery, Training Zones, and Stress):
Prompt: “I just came back from my long Z2 ride. Please assess it in detail against our training plan!”
1. SCORECARD vs PLAN
| Metric | Target | Actual | Grade |
|---|---|---|---|
| Duration | 120–150 min | 137.5 min | ✅ Perfect |
| Time in Z1+Z2 | — | 99.4% (130.1 min) | ✅ Excellent |
| Max HR | ≤130 (ceiling) | 135 bpm | ⚠️ Brief breach |
| HR drift | <5% | -4.5% | ✅ Excellent |
| Avg cadence | 85–95 rpm | 74 rpm | ❌ Low |
2. LAP-BY-LAP ANALYSIS
- Laps 3–6 were textbook Z2: HR locked at 121–122 bpm. This is exactly the metabolic zone we wanted.
- No cardiac drift: HR drift = -4.5%. This means your aerobic system handled the load comfortably.
安装
docker run -i --rm --network garmin-grafana_default --env-file /path/to/.env ghcr.io/ghighi3f/garmin-grafana-mcp-server:latest python server.py配置
{
"mcpServers": {
"garmin": {
"type": "sse",
"url": "http://<your-host>:8765/sse"
}
}
}