Home / Fabric sizer

One AI. The hardware you already own.

Every Augmentum node advertises what it can serve and borrows what it can't. Seven capability kinds route peer-to-peer between boxes you pair deliberately — and Fabric is off by default: a solo install never runs a line of it until you flip fabric_enabled.

Live routing

Seven capabilities. Cheapest capable peer wins.

Routing is cost-aware: each request goes to the cheapest peer advertising that capability. Click a peer to take it offline and watch its routes fail over — and watch what honestly can't be served without it.

When no online peer advertises a capability, the request stays local and reports no capable peer — Fabric borrows hardware, it doesn't invent it.

What federates

The seven kinds, exactly.

These are the capability kinds a node can advertise and a peer can borrow — no more, no less.

LLM inference

llama.cpp lane

Chat, agents, and coding turns. A weak device borrows a strong peer’s tokens.

Image generation

SD15 / SDXL / FLUX overlay

Diffusion jobs route to whichever box has the GPU headroom.

Text-to-speech

Kokoro · Pocket TTS

Sub-second replies rendered wherever a voice lane is warm.

Speech-to-text

Moonshine

Streaming transcription for ambient listening and dictation.

Knowledge search

packs + embeddings

Hybrid retrieval over your library, served by a peer that holds the index.

Code execution

sandboxed Python

Containerized interpreter runs on the box with cores to spare.

Trust model

Pairing is deliberate, not discovery.

No peer joins your fabric by being on the same network. Trust is established out-of-band, SSH-host-key-style, and every request rides an end-to-end encrypted envelope.

  1. 1

    Paste the fingerprint

    Each node shows its key fingerprint in the Fabric UI. On the initiating node, you paste the remote node's URL and fingerprint — the fingerprint is the out-of-band secret.

    SHA256:ab3f…9c2e
  2. 2

    Signed request

    The pair request is Ed25519-signed, carries a replay-blocking timestamp, and must address the right node's fingerprint — only the holder of the matching key, aimed at the right box, gets through.

    ed25519 ✓ · t+30s
  3. 3

    Approve on the far side

    The receiving node's operator approves the pair explicitly. Trust is never permanent — either side can revoke it at any time.

    peer approved

Three pairings, three jobs — deliberately separate schemes:

Fabric nodes your machines · URL + fingerprint paste, signed request, explicit approve
Connect other people · signed contact card as link, QR, or a short code, verified Signal-style with safety numbers
Cast receiver your TV · two-minute QR on the screen, approved by a phone that's logged in
— workloads in parallel
— devices in the fabric
— GB VRAM, combined
— faster than queued on one box
1

The hardware you have

Start from a preset, then tweak. Drag a chip from one device onto another to override.

2

What you want it to do

Each workload lands on the best-fitting device. Within LLM inference and Image generation, the largest selection wins.

LLM inference

Image generation

Text-to-speech

Speech-to-text

Knowledge search

Code execution

Cast rendering

3

What your fabric runs

This sketch allocates greedily — biggest workloads claim GPU first, CPU-capable ones fall back. The production router is cost-aware and prefers the cheapest capable peer.

Your fabric

One-box equivalent

Estimates use rough Q4_K_M VRAM plus KV cache headroom at ~8k context. Real numbers depend on quant, context length, model choice, and what else is on the GPU. Augmentum has the production allocator; this page is a planning sketch.

Side by side

Warm lanes beat model thrash.

Fabric keeps chat, image, voice, and memory models resident on the machines that serve them. One box has to queue work and pay load/swap costs as it jumps between modalities.

Fabric —

Each peer can keep its assigned modality warm: image on a GPU box, chat on another machine, voice and embeddings on light CPU lanes. Wall-clock is the slowest warm lane, not every model loading in sequence.

One box —

A single accelerator has to unload or page between chat, diffusion, voice, and embedding models. The amber bands are load/swap time; useful work waits behind them.

How placement affects speed

Why fabric

What you get for spreading the load.

Runs in parallel, not in a queue

One GPU serves one heavy workload at a time. An enabled fabric can run chat, image, voice, and embeddings on different peers so one job does not stall the whole house.

Reuses the hardware you already own

An old gaming PC, a Mac mini sitting on a shelf, a laptop you upgraded from — together they outclass a workstation card you'd have to buy.

Graceful when peers change

When a paired device is unavailable, Augmentum can fall back to the remaining advertised capabilities instead of assuming one perfect box is always online.

Stays on your configured network

Every peer is a machine you pair deliberately, usually on LAN or Tailscale. Requests travel in Ed25519-authenticated, end-to-end encrypted envelopes. There is no Augmentum-hosted relay, no metering, and no telemetry.

In practice

Three layouts Fabric is built for.

Voice while coding

Tower · Laptop

Tower runs the coder agent on the large model. Laptop handles Moonshine STT plus Kokoro or Pocket TTS for ambient voice. You dictate a code review request while the agent is mid-refactor; both can run at once.

Family of four

Tower · Mac mini · Old laptop

Four people, four conversations, four devices serving them. The tower handles the heaviest chat, the Mac mini runs voice for the kid, the old laptop does embeddings + memory writes for everyone.

Old hardware reuse

2020 desktop · Mac mini M2 · old laptop

Repurpose compatible machines you already own. The desktop handles image generation, the Mac mini runs a small chat model on CPU, and the old laptop takes embeddings or voice. Total spend: zero.