Put a model on the machine with the big GPU - a gaming PC, a Mac Studio, a DGX Spark - and chat with it from your laptop. No account, no cloud, no VPN. You approve each computer from one console, revoke it any time, and every answer carries a receipt signed by your own host.
The lines below are from the release acceptance run: a Windows helper with an RTX 4080 and a Linux host.
> mfab helper install-llama-serverThis computer: windows-x64, GPU flavour cuda.Installed llama-server for windows-x64 (cuda)> mfab setup --role helper --allow-peer-compute --model qwen2.5-1.5b-instruct-q4_k_m.gguf --join-url "http://<host>:8935/join?invite=..."Loading the cached model with llama-server (llama-server.exe)...Model ready in 8.7 sCompare this pairing phrase on the host: rearview graves cuddly campus tighten sixtieth
$ mfab host start --detachHelper join link: http://<host>:8935/join?invite=...# phrase matches - approved in the console$ mfab peers connect mani_mfabric_peer --key-file gateway.keyTrusted LAN only: this HTTP gateway does not encrypt prompts in transit.Connected. Chat with: mfab chat --peer mani_mfabric_peer "Hello"$ mfab chat --peer mani_mfabric_peer "Say hello in five words."Hello there!completion_tokens=4 runtime_route=openai_compatible_peer
The host is the computer you type on. A helper is a computer with a model and ideally a GPU. It serves only after you approve it on the host.
Install MFabric and start the host. It shows a join link for your other computers.
The helper installs the official llama.cpp build for its hardware and joins with the link. Check the pairing phrase matches, then approve.
The helper's GPU answers. Approvals survive a restart. Revoke a helper and it is refused from that moment on.
Each of these ran end to end on physical hardware, not just in unit tests. The Windows row used this exact release; the Mac row used a 0.2.0 pre-release build.
The signed installer on a laptop with an RTX 4080 served a model to a Linux host: pairing, approval, chat, signed receipt, host restart, revoke and re-approval all passed.
An Apple M4 Pro MacBook served the same model through llama.cpp's Metal build (0.2.0 pre-release), including answering after the host restarted and refusing after revocation.
Every remote answer carries a receipt signed by your own host, naming the approved computer and route. No prompt text, no key.
Downloads the official ggml-org release for your hardware, checks every file's published SHA-256, and swaps it in only when complete.
The host checks it can reach the helper and prints the exact fix: a narrow Windows Firewall rule, or the macOS Local Network permission.
Each helper shows a pairing phrase you compare on both screens before you approve it. Approvals persist across restarts; revoking refuses the helper immediately.
An NVIDIA DGX Spark (Linux, ARM64) was the host in both journeys, and the llama.cpp installer was verified on it. The same Python package installs on Windows, macOS and Linux.
Use the same version everywhere. MFabric does not update itself: to update, install the newer version the same way.
Download mfabric_consumer_setup_0.2.0.exe and run it. Windows should show the publisher as Sfiniti AI Inc. Because the signature is new, SmartScreen may show a caution first; choose "More info" and check the publisher before you continue. It bundles its own Python.
Download the wheel, check its fingerprint, then:
python3 -m venv ~/mfabric ~/mfabric/bin/pip install ./mfabric_consumer_public_beta-0.2.0-py3-none-any.whl source ~/mfabric/bin/activate mfab doctor
Mac helpers: when macOS asks whether Terminal may find devices on your local network, choose Allow (or System Settings > Privacy & Security > Local Network). Keep the Mac awake while it serves, for example with caffeinate -dimsu.
Download the wheel, check its fingerprint, then:
python3 -m venv ~/mfabric ~/mfabric/bin/pip install ./mfabric_consumer_public_beta-0.2.0-py3-none-any.whl source ~/mfabric/bin/activate mfab doctor
Linux helpers: if ufw or firewalld is active, allow the helper's gateway port (8788 by default) from your local network only.
# On the host (the computer you type on) mfab host start --detach # On the helper (the computer with the model) mfab helper install-llama-server mfab setup --role helper --allow-peer-compute --model <model.gguf> --gateway-key-file <key file> --join-url "<join link>" # On the host, after approving the helper (mfab host open) mfab peers connect <helper id> --key-file <key file> mfab chat --peer <helper id> "Hello"
Put the GGUF model file in your Downloads folder first; MFabric never downloads models on its own. The full step-by-step, including the gateway key file, is in the README inside the download.
All three are also in SHA256SUMS.txt.
# Windows PowerShell Get-FileHash -Algorithm SHA256 .\mfabric_consumer_setup_0.2.0.exe # macOS / Linux shasum -a 256 -c SHA256SUMS.txt
MFabric Consumer is licensed under the GNU AGPL v3.0. The complete source for this release is here, and the Windows installer carries the same source archive in its install folder.
Organizations that want to embed, modify or host MFabric without the AGPL obligations can get a commercial license from Sfiniti AI Inc.
Not yet. Traffic between the host and a helper is HTTP on your local network. A gateway key and your approval control who can use a helper, but they don't encrypt prompts in transit. Use it only on a network you trust, and never forward these ports on your router.
No prompts, answers or device data go to Sfiniti. Prompts and answers travel only between your own computers. The only outside connections are the ones you ask for: the llama.cpp release from GitHub, any model you fetch yourself, a frontier provider if you add your own key, and the manual "check for updates", which reads a static version file from sfiniti.ai.
No product telemetry and no automatic updates. Checking for updates only happens when you ask, and it sends nothing about your device. To update, download the newer version from this page and install it the same way.
Two or more computers on the same Wi-Fi or Ethernet network, a GGUF model file on the helper, and Python 3.11+ on macOS or Linux. A GPU on the helper makes answers fast; CPU-only works too, just slower.
Yes. It's open source under the AGPL. A commercial license is available for organizations that need different terms.
Email [email protected] with the subject "MFabric issue" and the output of mfab doctor --json. Never include a gateway key, join link or prompt text.
If it's not on this page, we're not claiming it.
Free, open source, and yours to run.