🧬 Harness Health Engineering

A Bryan-Johnson-grade health operating system you actually own. Your body's objective signals flow in automatically from Whoop; your subjective signals (mood, energy, social) flow in by hand; an on-device AI agent finds the patterns at the seam — all as plain Markdown on your machine.

The thesis

Quantified-self tools measure a lot and change almost nothing. They live in silos, they read your body but never your mind, and they hand you dashboards instead of decisions. Harness Health Engineering treats your health like a long-running engineering project:

• Whoop is your linter for the body — morning recovery surfaces the "error" before you compile it into overtraining or burnout.

• The repo is your memory — every day is durable, versioned, searchable evidence, not a number that scrolls off a feed.

• The agent is your synthesis layer — it correlates objective physiology with subjective state, the one place no wearable can reach.

The whole system is local-first: the only thing that ever leaves your machine is the outbound call to Whoop to pull your own numbers.

Related MCP server: MCP Index Notes

How it works

flowchart TD
    W["Whoop API v2"] -->|"OAuth2 · auto sync 9:00"| SYNC["scripts/whoop/sync.mjs"]
    H(["You · 20s/evening"]) -->|"mood · energy · supplements · social"| RAW
    SYNC -->|"recovery · HRV · sleep · strain"| RAW["01_raw · daily logs"]
    RAW --> SRC["02_sources · weekly notes<br/>FACT / INFERENCE"]
    SRC --> SYN["04_synthesis · patterns"]
    SYN --> EXP["05_decisions · experiments<br/>one variable at a time"]
    EXP --> OUT["06_outputs · plans · routines"]
    EXP -.->|"feedback → rule"| RAW

    RAG["on-device hybrid RAG<br/>e5-small · sqlite-vec · FTS5 · RRF"] --- KB[("Markdown KB 00→06")]
    MCP["MCP server<br/>kb_search · kb_think · kb_backlinks"] --- KB
    RAW --- KB
    SRC --- KB
    SYN --- KB

The cycle: Signal → Ingest → Source note → Synthesis → Decision/experiment → Output → Feedback → Iterate. Each loop makes you measurably smarter about yourself, instead of "I tried a thing once."

What it tracks (multi-dimensional, Blueprint-style)

Sleep & recovery · supplement stack + adherence · training load · activity types · nutrition timing · socialization · mental state (mood / energy / anxiety). Objective from Whoop, subjective by hand, value at the intersection.

Why it's different

Technology

See ARCHITECTURE.md for the deep dive.

Quickstart

corepack enable
pnpm run setup                 # semantic + skillopt + viewer sub-packages
cp .env.example .env           # add WHOOP_CLIENT_ID / WHOOP_CLIENT_SECRET
pnpm whoop:auth                # one-time OAuth (browser → .whoop/token.json, gitignored)
pnpm whoop:sync                # pull physiology into today's log
pnpm kb:index                  # build the semantic index (first run downloads ~120 MB ONNX)
pnpm kb:think "why did my mood drop this week?"
pnpm kb:doctor                 # KB health-check

Then the daily rhythm is: morning sync runs itself · evening log one line · Sunday ask the agent to synthesize the week.

Evening logging is easiest via the Telegram bot (scripts/bot/) — a free, zero-dependency, local "dumb collector" that drops whatever you text it into today's log. No LLM, no API key: the reasoning stays with the agent in Claude Code. See scripts/bot/README.md.

Privacy & safety

• Local-first. Your health data lives in this repo on your disk. Nothing is sent anywhere except the Whoop API call that fetches your own data.

• Secrets never committed. .env and .whoop/ (OAuth tokens) are gitignored and deny-read by the agent.

• Not a medical device. The agent never diagnoses or interprets symptoms as conditions. Any worrying signal resolves to a single recommendation: see a specialist.

Contributing

This started as one person's body-as-codebase — but the architecture generalizes to any self-quantifier. PRs welcome, especially: more wearable adapters (Oura, Garmin, Apple Health), a live Whoop MCP server, richer experiment templates, and synthesis evals. If you're building the future of human optimization and this resonates — open an issue, let's talk. (Yes, Bryan, this includes you.)

See CONTRIBUTING.md.

License

MIT © 2026 Christina Vinter. See LICENSE.