Connect each device once. The readings come to your own machine.
Ilfirin runs on your computer. It pulls the wearables and meters you already use straight to that machine, unions them into one local log, and you coach with it through a normal Claude chat no spreadsheets, no new app, no forms. Most of the day's signal is read for you; the rest takes seconds.
Coaching, not medical care — no longevity or anti-aging promises; the name is the long view, not a claim on your lifespan. Ilfirin runs on your own computer, reasoned by Claude (a frontier AI model) through the Claude plan you already pay for — so "no new subscription" means no extra fitness-app fee, not "free": you still need a Claude plan and an always-on machine. Privacy by architecture — no Ilfirin cloud, account, or database. Your records stay in files you own; during a chat, the readings Claude needs pass through Anthropic. An early, invite-based beta — not a finished product, and not pretending to be.
What runs where, stated plainly.
Three parts: the instruments you already own, a small program on your own computer, and you — in a Claude chat. The data path runs one direction. In, reasoned, out.
One connector, not a fleet. Claude.ai talks to Ilfirin through a single MCP connector — the one program on your machine that handles auth and routes every tool call. The wearable layer sits behind it: three direct-OAuth pull scripts — Google Health for Fitbit and Pixel, Withings for the scale and blood-pressure cuff, Garmin through a community client — write normalized records into local files, unioned at read time. No third-party data broker sits in the middle.
The wearable you already own — read for you.
Connect each device once. After that the pull is passive: a daily sync keeps your watch, scale, and cuff current with nothing to open or remember. Reads today: Fitbit · Pixel · Withings · Garmin (community) · Oura · Whoop · Apple Watch (via the free Health Auto Export app). Consumer CGM patches (Stelo, Lingo) are not yet — a Dexcom/Libre bridge can push glucose in. Stated, not implied.
Why Garmin reads "community." Garmin's official developer program is closed to small developers in practice, so the Garmin pull uses a maintained community client — best-effort, and disclosed as best-effort. Apple Watch comes in through a phone-side push app (the free Health Auto Export → Ilfirin's local endpoint), so it's supported with a one-time setup; consumer CGM patches (Stelo, Lingo) have no public API and return an honest "unavailable" — though a Dexcom/Libre bridge can push glucose in. Google Health replaces the sunsetting Fitbit Web API, so Fitbit and Pixel arrive through the same door.
Most of it is read for you. The rest takes seconds.
Most of the day-to-day signal — heart rate, sleep, steps, recovery — is captured automatically by the devices you already wear. The rest — food, bloodwork, and how you actually feel — you add in a normal chat. No forms either way.
| Reading | How it arrives | Cadence | Domain |
|---|---|---|---|
| VO₂max · steps · time in each HR zone | auto wearable | daily | Movement |
| Heart-rate recovery · HRV · resting HR vs baseline | auto wearable | daily | Movement · Recovery |
| Sleep duration · stages · regularity | auto wearable | nightly | Recovery |
| Body composition if owned | auto smart scale | per weigh-in | Fuel |
| Blood pressure if owned | auto connected cuff | per reading | Biology |
| Local PM2.5 if location set | auto EPA AirNow | hourly | Environment |
| Meals — a photo of the plate, or "a chicken bowl and an apple" | in chat photo or text | per meal · ~secs | Fuel |
| Bloodwork — drag the lab PDF in, markers extracted | in chat one drop | annual | Biology |
| Morning check-in — sleep, soreness, readiness | in chat | daily · ~30s | Recovery |
| Strength sets · quarterly benchmarks | in chat | workout days · qtrly | Movement |
| Environment check — connection, purpose, screenings | in chat | weekly · ~60s | Environment |
The dashboard is for reading. The chat is where you act.If you're filling out a form, the instrument failed.
All of it, in one local view — no single score.
Everything above lands in one glanceable dashboard served from your own machine at 127.0.0.1:8766. Five domains, the well-studied signals most apps skip, this week's plan. Read-only — every action happens in the chat. There is no composite "longevity score," because the components carry the information and a single number misleads. We say that out loud, as a virtue.
Focus: VO₂max up ~1.2 this month — hold the 4×4 cadence; don't add volume.
Focus: bedtime SD crept to 38 min; duration's fine — anchor a steady bedtime.
Focus: plant-protein share and oily fish are the two levers past 70.
Focus: 1 lab flag — LDL 121, above target. Discuss with your doctor.
Focus: 2 screenings overdue — book the annual physical this week.
Anonymized sample. Representative numbers, not anyone's real records. Every status carries a text label — never color alone. The headline is the single most-actionable signal across all five domains, never a composite figure.
What it takes to run, honestly.
Two honest paths today, and a third on the way. We tell you which is shipped and which is planned.
Self-host, today
Clone the repo, copy the profile template, run the first-run wizard. For a technical user that's about 60–90 minutes the first time terminal required. Claude.ai reaches your instance over a Cloudflare Tunnel with Cloudflare Access in front; or run it fully local against Claude Desktop with no tunnel at all. Either way it's one MCP connector.
- An always-on computer — Windows, Mac, or Linux — to hold your records and serve the dashboard.
- A Claude plan you already pay for. Ilfirin adds no AI fee of its own.
- The wearable accounts you already have — Fitbit / Google, Withings, Garmin.
Request a set-up instance
Tell us your devices and we'll provision your own instance and send a personalized installer with a guided setup — no command-line expertise needed. Access is limited while the beta stays small.
Desktop installer — planned
A bundled desktop installer (Python plus a wizard) is on the roadmap, targeting under 15 minutes with no terminal use. Planned, not shipped — flagged as planned, not promised.
It runs on your computer. Your records stay there.
Privacy here is by architecture, not by promise. There is no Ilfirin cloud, account, or database — nothing for us to sell or leak. The dashboard and your files live on your machine. During a chat, the readings Claude needs pass through Anthropic; we link their policy and don't pretend otherwise.
Best quality
All readings available to the coach during a chat. Best AI quality, best fit for most users.
Summaries only
Raw values stay local. Tools return ranges and trends — "within baseline" — instead of exact numbers.
Nothing leaves
Refuses the chat path entirely — nothing leaves your machine.
Every tool call is written to a local audit log you can read at any time — which tool ran and the shape of what it returned, never the values. Ask to see it anytime.
Want it set up for you?
Tell us about your devices and your machine. We'll provision your own instance and send a personalized installer with a guided setup — no command-line expertise needed; we walk you through every step. An early, invite-based beta; access is limited while we keep it small.
Coaching, not medical care — the name is the long view, not a claim on your lifespan. An early beta — not a finished product, and not pretending to be.