MFabric stitches the machines you own - a laptop, the box in the closet, the big machine
you and your friends bought together - into one private AI fabric. You bring your own keys; your prompts, files,
and models stay on your hardware. Chat with a model you already have on disk, approve
helper computers from one console, and check a privacy meter that shows exactly what stayed
on your machine and what went to your own provider - with a downloadable receipt that never
contains your key. The mfab command line is in the box too.
Anything experimental is clearly labeled and stays off until you turn it on.
Optional: release + security notes by email. No tracking pixels, never shared, and not tied to an in-product account or identity.
mfab CLI
See what changed →
The machines that run the biggest AI models cost thousands. You don't have to buy one alone. Chip in with friends, put the machine at one person's place, and everyone chats with it from their own computer — owned by your group, private to your group.
Compare: five people paying $20 a month for AI subscriptions spend $1,200 a year — and own nothing at the end. The same money buys the machine. It's yours, it gets better every time the open models get better, and nothing your group types ever leaves the machines you own.
The machine lives at one person's place. They run the host, load the models, and decide who's in.
Each member installs MFabric and pastes a join link. The owner clicks approve — and can remove anyone in one click. It's their machine.
Members talk to the big models from their own laptops. The machine takes turns, like a shared workstation — for a group chatting through the day, you'll rarely wait.
Straight talk: it's one conversation at a time, not a data center. And it's easiest today when the machines share a network — a house, an office, a dorm. Our simple across-town setup guide is next on the list.
Most "private AI" still puts a vendor between you and your data. MFabric keeps us out of that path entirely, asks before anything leaves your machines, and writes you a local receipt so you can check for yourself.
Bring your own key to any provider you like. We never hold it, see it, or proxy your calls - the request goes straight from your machine to your provider.
Prompts, files, and models live on hardware you control unless you explicitly choose your own provider. Storage helpers see sealed blocks; trusted compute helpers require separate approval because they may handle model-internal tensors. Nothing joins uninvited.
Normal work runs the proven way. The clever tricks - sharing memory across machines, splitting big models - are off until you turn them on, and they step aside the instant anything looks wrong.
The peer beta is a free download. Verify the signed installer or zip hash, join a host you control, and run on your own network. Optional onboarding is help, not a requirement.
Runs on your machine, gives right answers, and only uses helpers you've approved. This is what you get out of the box.
Bigger tricks for bigger jobs: keep long conversations alive across machines, split a model too big for one box, or send heavy work to a stronger machine. Each is off by default and labeled with exactly what it can and can't do.
Long or batch jobs go into a queue across your machines - they resume if interrupted, and you can pause or cancel them any time.
Install it, invite your other computers, approve them from one screen, load a model you already have, and chat. That core works today. The bigger tricks are in the box too - labeled as previews and experiments, because that's what they are.
Load a model from your own disk and talk to it - no account, no cloud. Words appear as they're generated when the model runs on your machine, and still stream when the answer comes from another machine you've approved. In the current published build, answers arrive complete instead; streaming ships with the next signed refresh.
Your laptop, a desktop, a rented GPU - MFabric moves work and data between them over encrypted connections that shut off rather than fail open. Apps built for the fabric use the pool; everything else just sees your normal machine.
Point MFabric at a stronger machine you're allowed to use - the desktop at home, a GPU you rent by the hour - and hand it the heavy job. Setting up a worker takes a few deliberate steps on purpose; this is not an open marketplace for strangers' computers.
The console shows where every answer went - what stayed on this machine and what left to
your own provider on your key - and lets you download a privacy receipt with no key inside.
Files stay private too: the assistant only sees folders you've explicitly granted. And for
the terminal crowd, the mfab command line is included.
Smaller models get a tuned engine that's measurably quicker; bigger models automatically use the standard engine. Faster where we measured it - not a universal speedup. The published installer still uses the standard engine for everything; the quicker one arrives with the next signed build.
The planner tells you whether your machines together can hold a model none of them can hold alone. Full split-model serving is an advanced preview - real, but not one-click yet.
Long chats normally push the oldest parts out of memory. This keeps them alive on your other machines instead of throwing them away. Today that path is slower than staying local, so it's opt-in and labeled a preview. We'd rather tell you than dress it up.
Run a small training test across your machines today. Full fine-tuning across helpers is a power-user lane; the goal is using hardware you already own, not a speed promise.
Every claim on this page is something you can check: a signature Windows verifies, a fingerprint you can compute, a receipt you can read. That discipline is the product.
mfab command line (run it with Python, no install needed). Your host still decides what each helper peer may do before it joins the fabric.You can download the peer today. Onboarding is still available for people who want help wiring a host, testing multiple machines, or running preview capacity paths.
The host holds policy, credentials, and model ownership, approves peers, and decides which jobs they may join. It is the center of your fabric.
A peer contributes bounded storage, workhorse execution, or training capacity - only what the host allows, revocable any time. On Windows, run the signed installer and choose Helper/peer; on macOS or Linux, use the peer package below.
Your host hands you a join link. That link is the secret handshake that connects a new device to your fabric. On Windows, the installer first asks whether this machine is the host/controller or a helper peer. Choose helper/peer to paste the join link. On macOS, Linux, or if you just like doing it by hand, the manual package is two commands.
Download mfabric_peer_setup_0.1.3.exe and run it. Before you click through,
make sure Windows shows the publisher as Sfiniti AI Inc. - that's how you know it's
really ours. The first screen asks Host/controller or Helper/peer. Choose Helper/peer to install this
device as a peer and paste your host join link. Choose Host/controller on your command machine to open the local host console.
Prefer to do it yourself? Unzip the peer package into a normal folder like
~/mfabric_public_beta, then run these two commands (drop your join link where it says so):
# macOS / Linux cd mfabric_public_beta python3 -m pip install -r requirements-agent.txt python3 mfabric_peer_bootstrap.py --join-url PASTE_HOST_JOIN_LINK --detach # Windows, the manual way (no installer) packaging\windows\START_MFABRIC_PEER.cmd
Back on your host dashboard, click approve. The host decides exactly what this device is allowed to do, and you can pause it or kick it off any time. You're always the one in charge.
A hash is just a fingerprint for a file. Run one command, and if the long code it prints back matches the one below, your download is genuine and untouched. The Windows installer is also signed, so Windows itself will show Sfiniti AI Inc. as the publisher.
# macOS / Linux shasum -a 256 mfabric_public_beta_peer_current.zip # Windows PowerShell - works for the installer too Get-FileHash -Algorithm SHA256 .\mfabric_peer_setup_0.1.3.exe Get-FileHash -Algorithm SHA256 .\mfabric_public_beta_peer_current.zip # Any platform with Python python -c "import hashlib,pathlib;print(hashlib.sha256(pathlib.Path('mfabric_peer_setup_0.1.3.exe').read_bytes()).hexdigest())"
We'd rather tell you than dress it up.
Nothing, unless you add your own key. If you choose to use an external provider, that request goes straight from your machine to your provider - MFabric never holds or proxies it. You can download the privacy receipt after any session; it contains no key and no prompt content, just a log of what stayed local and what route was taken. Check it yourself and see.
No. MFabric is free closed-source software in public beta. The manual beta package is not an open-source release and grants no open-source license. Trust comes from a signed publisher, published SHA-256 fingerprints, explicit outbound routes, local privacy receipts, and the ability to run without an Sfiniti account.
You shouldn't take our word for it. Check the SHA-256 hash against the fingerprint on this page before you run anything. Download the privacy receipt and read it - there is no key in it. The installer is signed by Sfiniti AI Inc. so Windows can verify the publisher independently. The receipts and hashes are the trust mechanism, not our assurances.
The published easy path is the signed Windows installer. The current macOS and Linux path is the manual beta package with Python 3.11 or later. You also need a compatible model already on disk or your own API key for an external provider. MFabric does not ship model weights.
MFabric sends no product telemetry. The update check runs only when you press the button and makes a standard HTTPS request for the public manifest; the web host may retain ordinary access and security logs. The email list is opt-in and is not tied to an in-product account or identity.
Advanced Preview features - like context retention - are receipt-backed and opt-in. They are not default promises and are not turned on unless you deliberately try them. Each one carries a receipt showing what was measured and under what conditions. "Advanced Preview" is an honest label, not a marketing upgrade.
Email [email protected] with the subject line "MFabric issue". Include the output of mfab doctor --json, or python mfab_cli.py doctor --json when using the current manual package. Never include an API key, join link, credential, private path, or prompt content.
Yes - that's the split-the-machine setup. One of you hosts it, approves each member, and can remove anyone in one click. It answers one conversation at a time, like a shared workstation, and it's easiest when the machines share a network today. Nothing the group types leaves the group's machines unless a member uses their own provider key.
Yes, the public beta is free. A business tier may come later; personal use stays free. No credit card, no access code, no account required to download and run it.
We keep this list honest on purpose. If it's not here, we're not claiming it.