AI Chat

Cadence includes a built-in AI coaching chat that uses your fitness data to provide personalized training advice.

How It Works

The chat is an agentic loop: when you ask a question, the LLM can call Cadence's MCP tools to fetch your data, analyze it, and respond with data-backed coaching advice. Each conversation supports up to 10 tool-calling rounds.

User: "Am I ready for a hard workout today?"
  ↓
LLM calls get_daily_brief → recovery 82%, HRV 65ms, 7.2h sleep
LLM calls get_readiness → green (training readiness: high)
LLM calls get_strength_history → last session 2 days ago
  ↓
AI: "You're in great shape for a hard session today. Recovery is 82%,
     HRV is above your baseline, and you've had 2 full rest days..."

Supported LLM Providers

Cadence uses LiteLLM to support multiple providers:

Provider	Model Examples	Env Variable
Anthropic	Claude Sonnet 4.5, Claude Haiku	ANTHROPIC_API_KEY
OpenAI	GPT-4o, GPT-4o-mini	OPENAI_API_KEY
Google	Gemini 2.0 Flash, Gemini Pro	GOOGLE_API_KEY
xAI	Grok	XAI_API_KEY
Groq	Llama, Mixtral	GROQ_API_KEY
Mistral	Mistral Large, Mistral Small	MISTRAL_API_KEY

API Key Precedence

1. Per-user key — users can set their own key in Settings
2. Admin key — stored encrypted in the database (set via admin panel)
3. Environment variable — fallback from .env

Coaching Personality

The AI operates as "Cadence", an expert fitness coach that:

• Leads with data — always calls get_daily_brief before giving advice
• Cites numbers — references actual recovery %, HRV in ms, sleep hours
• Uses readiness colors — green (go hard), yellow (moderate), red (rest/easy)
• Flags risks proactively — HRV decline 3+ days, sleep debt >2h, ACWR >1.3
• Renders charts — visualizes trends after fetching time-series data

Streaming

Chat responses are streamed via Server-Sent Events (SSE):

Event	Description
token	Text streamed token by token
tool_start	Tool execution beginning (shows loading state)
tool_end	Tool execution complete
chart	Chart spec for visualization
usage	Input/output token counts
done	Stream complete
error	Error occurred

Coaching Profile

Each user can set a coaching profile in Settings — free-text context about their goals, injuries, preferences, and training background. This is prepended to the system prompt for every conversation, personalizing the coaching advice.

Token Limits

Admins can set per-user monthly token limits via the admin panel. Token usage is tracked per conversation and per model in the token_usage table.