Dashboard restart needed. The page you are viewing is newer than the running dashboard process. Restart the dashboard so buttons and APIs match this UI. python mfabric_dashboard.py

mfabric.

private AI on the computers you already own

MFabric turns the computers you already own into a private AI network. Run local models, approve helper machines, and keep sensitive work on your own hardware — with a privacy receipt to prove it. No cloud unless you bring your own key.

Updates

Manual check only — does not upload your device list, model list, or prompts.
gateway offline no model single-stream
0
whole fabric

What MFabric can do

loading capability map
Use nowThese work today without special setup.
Advanced betaWorks, but needs setup or owner approval.
Previews & experimentsPreviews are receipt-backed and opt-in; experiments are rough.
This console starts with chat and task batching because those are the fastest things to try. Click any card above to see its controls, or use the Show me around button for a first-run walkthrough. The cards below show the full beta surface and where each capability lives today.
1
the model

Load an open-weight model

Open this console from the mfabric agent to load a model.
↑ Pick a model above and press  Load  to enable the swarm and chat below. No local model yet? You can also chat right away with just your own API key — see Use an outside model further down.
2
save model work

Smart answer routing

where answers came from
Each task starts here. Easy tasks can answer from memory or cache. Hard tasks wake the loaded model. When batch tasks is on, multiple hard tasks can run together instead of one at a time.
3
your work

Run several tasks

one task per line
load a model first
No agents yet. Load a model, then run a swarm (or hit "Load example").
mfabric swarm grid
=
live · gpu

Telemetry

GPU memory per machine, sampled live. Bounded prefill/decode holds the curve flat as context and agents grow — a dense run climbs until it OOMs.
+
one-off

Single chat

load a model first
Talk to the model directly.
F
file access

Allowed folders

agent file access consent

Choose folders this console's agent can use. Nothing is available until you pick it here, and you can remove access at any time.

Pick a folder with the Browse button; the assistant can only ever see folders you picked, and you can remove one any time.

No folder picker opened yet.
Advanced: paste a folder path instead
No folders allowed yet. Choose a folder to let the agent use files inside it.
Recent file access
No files accessed yet. Choose a folder above, then ask the agent to use files from that folder.
File access stays limited to folders you approve here: hidden files, symlinks outside allowed folders, and path traversal are rejected. Writes are UTF-8 text, local-admin-only, and overwrite-protected unless requested. The audit log stores filenames only.
→
optional

Use an outside model (your key)

Your key stays on this machine. It is sent only to the local mfabric agent, which calls the provider directly from here — never through an sfiniti server, never logged.
T
experimental beta

Train on spare devices

bounded public/synthetic adapter jobs
Step 1 — Connect your training peer
Paste the relay URL of the peer that will run the bounded adapter job.
Show all fields
Experimental — ready to configure a bounded training job.
This runs a synthetic shaped test — it does not touch your data or save a real adapter. Experimental means visible and usable for approved test jobs, not product-default training. This panel sends only bounded synthetic or public-activation adapter work through the signed relay. Private prompts, private corpora, model weights, shell commands, and internal private payloads stay blocked.
*
advanced beta

Use a stronger GPU

owner approval required
Step 1 — Connect your GPU machine
Paste this host's join link (printed in the terminal when you started the host). It looks like http://.../join?invite=...
Show all fields
Idle — relay secret stays server-side.
Join idle.
Use this only after you have an approved desktop, rented GPU, or trusted friend's machine ready. This panel sends bounded test jobs only. Private prompts, private corpora, model weights, shell commands, and raw relay secrets are not accepted here.
~
secondary · the fabric

Connected computers

Privacy: MFabric only sees devices already running MFabric on your network. They cannot join unless you invite them and they accept.
Any other computer on your network running mfabric joins this pool on its own and shows up in the swarm’s run on picker — pick a higher-powered box to execute a swarm or chat there while you keep working on this one, and open this dashboard’s session link on either computer to watch the same run. Optional: everything above works on this machine alone.
Two ways to add a machine. Automatic: on the other computer (same network) start the mfabric agent — it announces over your LAN and appears under Connected computers within seconds. Manual: if you already know its address, add it here — it’s checked for reachability and then shows in the swarm’s run on picker.
How it works. One open-weight model loads once and is shared by every task. MFabric answers what it can from local memory or approved caches and only wakes the loaded model when it must; batch mode groups hard tasks so the GPU can work on several at once.
Honest boundary. Each line here is one routed task, not yet a tool-using multi-step agent loop. Batch mode is the throughput win today; the memory ceiling for many long-running tasks is still gated work and is not claimed here.
Local memory. Answers you generate are remembered on THIS machine only (folder runtime_durable_memory; question keys and your identity are stored hashed) so repeats answer instantly. Nothing syncs anywhere.