nest

AI fitness tracker where the model never touches a number you act on.

Nest is a personal fitness tracker installed to the iPhone home screen as a PWA. An AI reads meal photos, answers questions about your own log, and writes coaching notes, but it is never allowed to set, suggest, or adjust a target. That line is enforced three ways: the structured-output schemas have no field for a target, a guardrail rejects any energy or macro figure the model was not shown, and the MCP tools offer no way to change a goal. The targets themselves come from pure, unit-tested TypeScript, including an adaptive TDEE that re-estimates maintenance from logged intake and the weight trend.

technologies

• next.js
• react
• typescript
• supabase
• openai api
• model context protocol
• zod
• vitest

key highlights

• Numeric guardrail rejects any figure not in the model's context, retries once, then falls back to fixed text
• Adaptive TDEE clamped to ±15% of the formula, at most 5% per update, and never below BMR
• Meal estimates stored as ranges, counted at the bound that keeps the goal honest
• MCP server with 13 tools, OAuth with mandatory PKCE, and an audit log shown in settings
• Offline-first logging through an IndexedDB queue that replays idempotently
• Around 320 tests covering nutrition maths, timezone edges, the guardrail and the OAuth flow