Macula Mesh Use Build Macula for… Docs Contact

Macula · sectors · All Macula for… pages

Macula for Running AI

Macula lets you run AI models on your own GPUs, or on GPUs another member of your realm shares, and reach them from anywhere without opening a port. A small bridge beside your NVIDIA PAIR cluster makes only outbound connections to the Macula Mesh and answers members of the realm it joined, and no one else. By default that is every member whose node is bound to a person who confirmed an email address; in Macula's community realm, that is anyone who has signed up and confirmed one. A hosted AI API, by contrast, receives every prompt you send and bills you per token. We are looking for partners to build it with.

Without Macula: a hosted API sees every prompt, or a home GPU server needs an open port

Plenty of homes and small offices already own a capable graphics card. NVIDIA PAIR, the Personal AI Router, turns those machines into one local inference cluster. It answers only on the local network by design, so using it from outside the house needs something extra.

prompts leave home

A hosted AI provider receives every prompt

A hosted model sees what you ask and what it answers, and bills you per token. Your notes, your code and your family's questions are sent to the provider's servers.

open port or whole network

A port forward opens the cluster to everyone, and a VPN opens the whole home network

A port forward is open to everyone on the internet. A VPN lets a device onto your whole home network, not just the model.

admin calls exposed

An inference server open to the internet exposes its admin calls too

An inference server that the internet can reach also offers downloading models, deleting them and controlling the engines. Whoever reaches one reaches all of them.

With Macula: a bridge that dials out and answers only members of the realm it joined

The Macula Mesh is a network of stations. Stations link up to form the mesh, keep its phonebook (a Kademlia DHT: which node offers what, and where to reach it), and relay calls between the parties a realm admits. There is one mesh and many Macula Realms. Each realm is an identity space on the mesh with its own key: its operator admits anything that can be given an identity, such as people, organisations, services, machines and devices, gives each a certificate signed with that key, and can revoke it. Anyone can run a station, as many as they like; the more stations, the stronger the mesh. Macula runs a number of seed stations and a community realm; a company or sector body can run its own realm under a Macula licence. The bridge then decides which of them may call it: by default any member whose node is bound to a person who confirmed an email address. The member's app and the bridge beside your cluster only make outbound connections; neither accepts an incoming one. Stations link to each other, so a member on the road reaches a bridge at home that dialled a different station.

The member's app and your bridge each make an outbound connection to a station, and the linked stations carry the call between them. The prompt is encrypted to your bridge, so the stations cannot read it. The bridge hands the request to PAIR on the same machine, PAIR picks a GPU in your cluster, and the answer goes back the same way. Your prompt runs on your hardware, never on an AI provider's.

What Macula promises

no inbound port

Nothing to forward on your router

The bridge connects out to the mesh, and PAIR keeps answering only its own machine. You retire the port forward and the VPN.

members only

A caller outside your realm is refused before any work starts

Every call carries a certificate your realm signed. The bridge accepts by default any member whose node is bound to a person who confirmed an email address, and refuses a machine that joined with no confirmed person behind it. The mesh checks it before the call reaches the bridge.

one operation

Chat in, answer out, nothing else

The bridge builds each request to PAIR itself and never passes on a caller's path. Downloading models, deleting them and engine control cannot be reached from your realm.

rate limit

Each calling device is rate-limited

Each caller gets a number of calls per time window that you set to what your hardware can carry. A runaway loop hits the limit and is refused.

revocable

Membership can be ended

Unbind one of a member's devices, or have your realm's admin end the member altogether. Membership tokens are short-lived, so a removed member is refused once the last token expires.

PAIR untouched

Your cluster keeps working as it does

No fork and no patch. PAIR still spreads work across your machines and fails over between them, and the programs on your own network keep using it as before.

Where it fits

Any group that wants to share GPUs it owns. Members reach the cluster through a Macula client that calls the bridge's chat procedure; the machines inside the cluster need no change at all. In Macula's community realm every confirmed member can call the bridge; a group that wants only people it knows runs its own realm and admits them.

A family's gaming PC, sharedA small office's GPU box, for staff on the roadA club or association pooling its machinesA research group's workstationPrivate assistants that keep prompts at home

Build it with us

Partners bring the clusters, the members and the applications; we bring the station software, seed stations, the bridge to NVIDIA PAIR, which is open source, and a place in Macula's community realm or a realm of your own under a Macula licence. We are looking for partners in three roles.

bridge a cluster

Households and small offices with GPUs

You own an NVIDIA PAIR cluster, or the cards for one. Bridge it into Macula's community realm or a realm of your own, and share it with that realm's members.

build a client

App builders

You build an assistant, a tool or an app that should send its prompts to hardware its users trust. Build it as a Macula client that calls the bridge.

pilot or research

Clubs, associations and research groups

A group that wants to pool its machines for its own members, run shared clusters with us, and tell us what it needs.

Run the bridge

You need Erlang/OTP 28 with rebar3 and a Rust toolchain. The tests stand in for PAIR and the mesh, so they run without either.

# build and test git clone https://github.com/macula-services/mcl-nvidia-pair cd mcl-nvidia-pair rebar3 compile rebar3 eunit

To bridge a real cluster, run it on a machine that is already a PAIR member, with your realm and a station to dial. Getting started lists every setting and how a member is admitted.