domestique-ai
DomestiqueAI is a self-hosted cycling training companion powered by a local LLM that never makes up numbers. It ingests rides and recovery data, computes metrics like CTL/ATL/TSB and hr-TSS, flags overtraining before symptoms appear, and lets you talk to a coach grounded in your real data.
Ficha actualizada 23 sept 2026, 17:09 · upstream 23 sept 2026, 16:57 · release sin release detectada
Qué es domestique-ai
Descripción editorial construida a partir del repositorio, su documentación y sus señales de mantenimiento.
DomestiqueAI is the opposite of tools like TrainingPeaks — it's not expensive, closed, or owned by a third party. It gives you full ownership: SQLite runs on your own hardware with no cloud subscription or data broker. The core innovation is an agentic LLM coach that can't lie about numbers; every metric comes from tested Python code rather than model hallucination. You own the data, run it locally, and get a coach that actually knows what's happening because its reasoning traces back to verified tool calls.
- Self-managed training planning with no reliance on paid platforms
- Overtraining detection before you feel symptoms (fatigue, soreness)
- Comprehensive recovery monitoring via HRV and sleep data
- Training load visualization using CTL/ATL/TSB metrics
- Personalized coaching conversations that cite your actual numbers
- Offline-capable training with GPS maps and activity export
- A real sports-science engine: hr-TSS is a Banister exponential TRIMP normalized to exactly 100 points at threshold, making it interchangeable with power-based TSS
- An agentic coach with guardrails that never produces unsourced numbers; the LLM only selects choices and the code computes everything else
- Overtraining detection using four physiological indicators (chronic TSB, Monotony, Strain, weekly volume jump) plus a health module tracking HRV, resting heart rate, sleep stages, SpO2, and skin temperature
- A privacy-first architecture: SQLite on your own hardware with no data leaving the device; all API calls are local tool invocations
- An installable PWA with offline support, live charts, GPX export, and webcal subscription feeds for calendar integration
- Requires a minimum of 4 CPU cores and 8GB RAM to run the full stack on a Raspberry Pi 5
- Garmin Connect sync is only supported as the sole source of activity data (no alternative sources)
- The Google Health API integration requires an active Fitbit or Pixel Watch subscription for HRV, SpO2, and sleep data
- The LLM-based coach may produce occasional hallucinated explanations if tool calls fail silently
- No built-in recovery prescription; the system detects overtraining but does not prescribe specific recovery protocols
Capturas del proyecto
Imágenes reales detectadas en la documentación del repositorio.
Perfil de infraestructura
Estimación inicial derivada del repositorio y de los servicios detectados.
Despliegue detectado
Qué señales reales de instalación hemos encontrado.
Run on a Raspberry Pi 5 with Docker Compose. Use Tailscale for local network access from anywhere. The .env file contains all configuration (FTP paths, database URLs, API keys). For production, consider adding healthcheck monitoring and alerting via Pushover or a dedicated monitoring service. The architecture is designed to be easily extendable — add more tool types or data sources by following the existing pattern of typed tools with type-safe interfaces.
Stack y dependencias
Componentes inferidos a partir de manifests, configuración y documentación.
Puede sustituir a
Relaciones detectadas por el análisis del repositorio y su categoría.
Experiencia real de la comunidad
Ayuda a contrastar la estimación automática con instalaciones reales.
Todavía no hay datos comunitarios para este proyecto.