Project story
From SaaS to AI agent
The story behind this rebuild lives on Substack. It continues the journey told in From cycling to the web.
Training copilot · AI-native · Open source
In 2018 I launched a training SaaS. In 2026 I unplugged it and rebuilt it AI-native: zero server, zero account — it runs on your own AI. It reads your day — load × feel — and tells you green, amber or red, and above all why. It reads; you decide.
The problem
First trap: you listen to the numbers. Fresh form, low load, the app says “go” — and you push through a day your body was clearly refusing. A week later, you pay for it.
Second trap: you listen to your morning mood, and you miss the days you were genuinely ready to push.
“We program the load. We forget the state of the person absorbing it.”
The method — open on purpose
The real subject is neither the numbers nor the feel alone. It's the crossing of the two — and knowing, in doubt, which one wins.
From your activities, feezify computes your load as true exponential averages (TSS → CTL / ATL / TSB). TSB gives your form. It's deterministic — math, not an opinion.
Your journal gives your real readiness: sleep, fatigue, motivation, stress, how the legs feel, what's happening outside sport.
feezify sets a light — green, amber or red — and, above all, the explanation. The colour isn't up for debate; the why is in plain words.
The rule — the one that carried all the value of the old app: the subjective gates the objective.
Green — go.
Readiness high, form non-negative. Both sides agree: today can take quality.
Let's be plain: feezify is a copilot, a reading assistant. It prescribes no workout, never tells you what to do, replaces neither a trainer nor a doctor. Weight, for instance, is only read as a trend — never as a diagnosis. The decision is yours, always.
AI-native · zero infra
feezify is no longer an app I host. It's a skill + a markdown core you own, installed on your AI (Claude, OpenClaw). The old web app maps tier for tier onto the new agent:
| Classic web app | The tier | AI-native feezify |
|---|---|---|
| Browser / frontend | interface | The LLM / the chatbot — Claude, OpenClaw |
| Server / backend | engine | The agent — a pure, unit-tested method core + adapters (Strava…) |
| Database | persistence | A second-brain memory — your journal + the copilot's compiled memory of you |
A portable folder of markdown you own. No server, no account, no telemetry. A connector (Strava…) stays off until you switch it on.
Your journal, which you write; and the memory the copilot compiles about you over time — every claim traceable to a precise day.
Strava is just an adapter; the engine depends on no brand. The code and the method are in the open — read them, judge them, fork them.
Project story
The story behind this rebuild lives on Substack. It continues the journey told in From cycling to the web.
Who it's for
You train seriously — bike, run, endurance — and you record your sessions.
You keep — or you're ready to keep — an honest journal, even a short one.
You want to read your form more clearly, not to be told your workouts.
You're comfortable with an AI (Claude, OpenClaw) and the idea of a tool with no app, no account.
Install
The no-terminal zip was never published. What works is the developer path: a terminal, Node and pnpm. Same engine on Claude and OpenClaw.
git clone https://github.com/jn-prod/feezify
cd feezify
pnpm install && pnpm buildcp -r core-template ~/my-feezify-coreYour private folder: baselines, sport, journal. Strava stays optional: your tokens go in .env, and feezify can read your day from your journal alone.
Point your AI at the skill's SKILL.md (Claude or OpenClaw), then ask: “How am I today?”
Documentation
The full manual ships with the site — the idea, the method, the install, privacy. Version française
The idea, what it does and does not do, and how a SaaS became an agent.
The developer path, step by step.
Your core folder, how to journal, how to ask for — and read — the read.
TSB, readiness, and the crossing rule — open on purpose.
Local-first, consent-gated connectors, and how to revoke everything.
Common questions and fixes.
FAQ
No, and it's deliberate. feezify reads your day and tells you green / amber / red, with the why. The decision stays yours. It's a copilot: it illuminates, you decide.
No. There's no app to host, no account to create. feezify is a skill + a markdown folder running on your own AI. Zero infra.
Home. Your core is a folder of markdown on your machine. No feezify server, no telemetry. A connector like Strava stays off until you switch it on yourself.
An AI that supports skills — Claude, or a self-hosted agent like OpenClaw. You install feezify on it, and ask it how you are today.
Today, yes: enough to use a terminal. The install means cloning the repo and building it. The no-terminal path, a zip to upload in Claude, was never published.
Activities come in through Strava, the only adapter built. The engine is provider-agnostic by design, so a watch provider would be an adapter, not a rewrite. Sleep, fatigue and feel enter through your journal — your watch's numbers can inform what you write, and an honest line about a bad night often reads truer than a sleep score.
Strava gives you numbers. feezify crosses them with how you feel, applies a rule (the subjective gates), and returns an explained read — not one more table to interpret alone.
The core works and the repo installs. I stopped developing it in September 2026, so don't expect new features. The code stays open: read it, judge it, fork it.
I stopped developing it in September 2026. A tool that reads your day — no server, no account, no orders — stays open: read the code, fork it.