Run a bigger model than any one of your machines can hold.
ModelCoor turns the laptop, the gaming PC, the old Mac and the phones in the drawer into one local AI — over the Wi‑Fi you already have — and lets you use it from anywhere. Private. Yours. Useful first, faster with every device you add.
Start free
Free for 3 devices · no card · macOS, Windows, Linux, Android, iOS
Why it works when other local AI doesn't
Every device counts
Big machines take whole layers; a 500 MB phone takes individual tensors. Placement is by measured free memory, not guesswork.
Nothing to configure
Devices find each other on your network by themselves, wired or Wi‑Fi. Every timeout assumes slow, flaky hardware, because yours is.
It keeps answering
A small model replies while the big one loads. Unplug a phone and the fleet re-plans around it. No babysitting.
Works with what you use
Speaks OpenAI, Ollama and Anthropic APIs — chat apps, IDE plugins and scripts built for any of them point at your fleet unchanged. Vision models answer about images too.
Reach it from anywhere
Every account gets a hosted URL to its own fleet — from the train, without opening a port at home.
You can see the whole fleet
Which device holds what, tokens served, time online — per device, in the dashboard and the mobile app.
What's actually running — and where your data lives
1 · The coordinator (your machine)
One process on whichever computer holds the model file. It plans the split, serves the API, and runs the dashboard where you pick models — with honest numbers for fit, speed and context before you download. Your prompts, chats and models live here and on your devices. Full stop.
2 · Nodes (everything else)
Every other computer and phone runs a node. On your network it finds the coordinator by itself and lends memory; the iPhone/iPad app is a node too, plus a fleet view you can check from anywhere. Each node downloads only its own slice of the model, straight from the coordinator or the public source.
3 · Our cloud (this site)
Accounts live here — and only accounts: your login, devices list, license, usage counters. It issues your license (verified offline, so your fleet keeps working if we're down) and, when you're away from home, relays your requests to your own coordinator without you opening a port. We never store prompts, outputs or models — relayed traffic passes through in flight and is gone.
AI traffic — weights, tokens, prompts — flows between your machines on your network. The cloud carries sign-in, licensing and (on hosted access) requests in transit. That's the whole deal; the privacy policy says the same thing in longer sentences.
Three steps to running
- Install it on the machine that will hold the model, and pull one:
backbone-lite models pull qwen2.5-7b
- Start it and link your account.
backbone-lite up — then paste the link command from your account page. That machine becomes the coordinator.
- Add every other device.
backbone-lite up on any other computer; it finds the coordinator by itself. On phones, install the app and tap Start.
Backends: Metal, CUDA, Vulkan, SYCL (Intel Arc), CPU, Android and iOS. Everything measured, everything configurable, and nothing leaves your network unless you ask it to.
Pricing
Free is a real plan, not a trial. Paid adds devices and lifts the hosted-access cap.
Free
$0 — 3 devices, 1 fleet, 1,000 hosted requests a day.
Home Pro
$9/mo — 12 devices, unlimited hosted access to your own fleet.
Studio
$29/mo — 20 devices, up to 3 sites, priority support.
Start free Compare plans