Technical overview

Your hardware. Your model.
Your rules.

Biolink runs as a small Docker stack on a machine you own. Health-data inference is local by default; frontier AI connects through your own scoped, revocable MCP link — never brokered by us. Switching backends is one config change.

4
PLATFORMS
3
INFERENCE TIERS
2
ARCHITECTURES
8+1
DATA SOURCES
1
PUBLISHED PORT
Supported platforms

From a Pi in a drawer to a DGX Spark on your desk.

One installer detects your platform and applies the right overlay. Multi-arch images — ARM64 and x86_64 — built locally from source. Nothing phones home.

HARDWAREOPERATING SYSTEMARCHPROFILENOTES
DEDICATED NODE — ALWAYS-ON, RECOMMENDED
{{ p.hw }} {{ p.os }} {{ p.arch }} {{ p.profile }} {{ p.notes }}
EVALUATION — YOUR EVERYDAY MACHINE, VIA DOCKER DESKTOP
{{ p.hw }} {{ p.os }} {{ p.arch }} {{ p.profile }} {{ p.notes }}
Same containers, same features everywhere — the profile only tunes memory, logging, and inference defaults. Evaluate on your laptop, then migrate to a dedicated node with one backup + one restore. Fully working on evaluation machines; an always-on box keeps overnight reports and scheduled syncs running.
Inference, your way

Local by default. Frontier by your choice.

Health-data inference runs on hardware you own. Frontier models connect the other way around — your own Claude or ChatGPT reads through a scoped MCP link you control. Pick a tier to see the exact config; with no model configured, everything except chat and narrative reports still works.

LOCAL — SAME MACHINE
PRIVATE — HARDWARE ON YOUR NETWORK
FRONTIER — YOUR OWN AI, VIA MCP
{{ selFile }} {{ sel.tierLabel }}
# {{ sel.comment }}
BIOLINK_AI_PROVIDER_PRESET={{ sel.preset }}
BIOLINK_AI_BASE_URL={{ sel.base }}
BIOLINK_AI_MODEL_NAME={{ sel.model }}
$ node/scripts/restart.sh
$ node/scripts/doctor.sh # probes the endpoint
clientClaude Desktop · any MCP client
authOAuth 2.1 · tokens you issue
scoperead-only · per-domain · you grant
auditevery tool call logged
revokeone click — access dies immediately
{{ sel.egressTitle }} — {{ sel.egress }}
CONFIG-ONLY SWITCHING
Backends are profiles (profiles/*.env). Change one file, restart, done.
PARITY HARNESS
A preflight + parity suite proves chat behaves identically across every backend.
GRACEFUL WITHOUT A MODEL
No GPU? Dashboards, sync, nutrition, and data reports all run model-free.
Install

One command. Safe to re-run, forever.

The installer detects your platform, installs Docker if missing, generates strong per-install secrets, builds from source, verifies health, and prints your pairing instructions. Re-running never touches existing config or data.

~5 min
FIRST BUILD · X86
15–30 min
FIRST BUILD · PI 5
--dry-run
PREVIEW MODE
install
$ curl -fsSL …/install.sh | bash
→ platform detected raspberry pi 5 (arm64)
→ docker + compose installed
→ per-install secrets generated
→ stack built & started healthy
→ smoke test passed ✓
Web UI → http://192.168.1.50:8787/app
Pairing → http://192.168.1.50:8787/pair/qr
Day-2 operations

Appliance-grade. One script each.

Every failure the doctor finds prints its exact fix. Updates back up first and roll back on failure. Restores keep your phone paired.

{{ op.cmd }} {{ op.desc }}
DATA SOURCES
Eight wearable clouds, official APIs.

Plus phone health data pushed straight from the app — no export files. Sources merge with priority arbitration so nothing double-counts.

OURA WHOOP GARMIN FITBIT POLAR SUUNTO STRAVA ULTRAHUMAN PHONE HEALTH
OPENS UP, ON YOUR TERMS
LAN-only by default. Wider by consent.
DEFAULTOne published port (:8787) on your LAN — web UI, admin console, pairing.
MCPClaude Desktop and other MCP clients read through scoped OAuth 2.1 tokens — read-only, audited, revoked in one click.
REMOTEOnly the AI-connection surface is ever exposed — device pairing and admin stay local-only.
BACKUPOne portable archive — keys and pairing state included. Restore on any machine.

Pick a machine. Run one command.

Free to run on your own hardware. Everything builds from source on your machine.