OpenPotluck.

Local AI, better together

Bring one dish.
Enjoy the
whole table.

Your machine has a specialty.
Your friends’ machines have others.
What if everyone could share?

Meet OpenPotluck: one local gateway for your AI models, with a table of trusted machines for everything else.

A place for every machine

In development. A first look at what we’re building.

A miniature communal table with three different computers, sharing sculptures that represent thought, images, and sound.
FIG. 01 Different machines. A much bigger menu.
One local gateway ChatCodeListenSpeakCreate

01 / The bigger picture

A great machine.
An even better table.

You shouldn’t need a room full of hardware to have a full set of AI tools. Keep what your machine does best close at hand. Let the table fill in the rest.

The Potluck GatewayAn illustrated example
Your toolsEditorChatScript
On your machineYour basketOne familiar local endpoint
Your dishReasoningReady on your MacLocal comes first
A friend’s dishImages & speechReady on their GPUThere when you need it

Your hardware. Your first priority.

The plan starts at home: your own models answer locally, without spending servings. You decide how much spare capacity to offer, and your work always comes first.

More possibilities. Less model swapping.

A memory-rich Mac can keep a reasoning model ready. A friend’s NVIDIA workstation can bring images and speech. The gateway routes a whole request to the machine that has the right dish hot.

A little give. A little take.

Serve a friend with your idle capacity and earn servings. Spend them on another dish, or buy them when you need more. Servings are service credits, never cash payouts or crypto.

The architecture / A closer look

Your machine stays yours.
The table connects the rest.

Two members bring hardware. Another brings only an idea. The same table connects all three.

Three members, one OpenPotluck table Member one runs a coding harness and a basket with language and speech dishes. Member two runs a coding harness and a basket with an image dish. Each basket's Potluck Gateway serves local requests inside its machine and dials outward to the table through one encrypted stream for shared work. A third member has no basket or AI hardware and uses the hosted endpoint with purchased servings. All shared routes go through the table, never directly between baskets. Answers return along the same connections. MEMBER ONE / THEIR MACHINE Coding harness Basket Potluck Gateway · local first Dishes: language + speech MEMBER TWO / THEIR MACHINE Coding harness Basket Potluck Gateway · local first Dish: image generation MEMBER THREE / JUST THEIR TOOL No AI hardware required. Harness → hosted endpoint No basket. Uses bought servings. A seat at the same table. OPENPOTLUCK The table Dishes · routing Servings ledger Baskets dial out. Nobody opens ports. Shared requests go through the table. Answers return on the same connections.
FIG. 02Planned architecture. Baskets dial out over an encrypted stream; nobody opens ports. Local dishes answer inside their own machine. Shared requests and replies pass through the table.

02 / Less tinkering, more doing

Good hardware
deserves a
great recipe.

Model tuning is a rabbit hole. The cookbook is our way out: curated, signed recipes matched to your machine’s memory, accelerator, and available resources.

A recipe pins the model, quantization, engine, and settings together. Speed and quality sit beside each other. You see what you’re running, and why it fits.

A few words from our kitchen
From the cookbook✳

RECIPE / 001 · CONCEPT

A little more
room to think.

A reasoning dish for a memory-rich Mac.

Hardware
Apple Silicon
Model & quant
Named, pinned, visible
Engine & settings
Matched to the machine
Evidence
Speed + quality + memory
Curated. Benchmarked. Signed.The recipe standard we’re building toward.

Anatomy of a recipe. Not a performance claim.

03 / Everyone brings something

Different strengths.
Same table.

Starting with macOS and Linux. From Apple Silicon and NVIDIA GPUs to roomy unified-memory systems, the cookbook is designed around what you actually have.

The deep thinker

Room for big ideas.

A Mac Studio’s unified memory makes space for larger language and reasoning models.

Apple Silicon · native engines
The hands-on creator

A gift for making.

A Linux GPU box can bring image generation, a coding model, or a fast speech engine to the table.

NVIDIA · images, code & audio
The next possibility

More kinds of capable.

DGX and AMD unified-memory machines expand the menu. Recipes will be qualified hardware by hardware.

A growing cookbook

04 / The people at your table matter

Sharing begins
with trust.

A friend’s machine still sees the request it serves. That’s why private tables are the starting point, and your own machine is always the first choice.

01

Your machine, your rules.

Choose the capacity, schedule, and tables you share with. Your basket enforces those limits locally.

02

Know what’s on the table.

Signed recipes protect the software an honest basket installs. Planned external checks help detect misconfigured or dishonest peers; signatures alone can’t prove how a peer behaves.

03

Measure performance, not conversations.

Table participation requires speed, latency, memory, and recipe telemetry. Prompts and outputs are not part of that telemetry.

04

Open by design.

The basket and gateway are being built to be open source. A public table is a later chapter, with more work to do on verification and privacy.

A small field guide

Welcome to the vocabulary.

A few familiar words for a different way to think about AI infrastructure.

Basket
The software on one machine. It runs your dishes and exposes your local gateway.
Dish
A model instance, running a specific recipe within a card or memory budget.
Table
A group of baskets, with its own members, routing, and servings ledger.
Cookbook
The collection of curated recipes, organized around real hardware.
Recipe
A pinned model build, engine, and settings, backed by benchmark evidence.
Serving
A credit earned by serving, or bought, then spent on someone else’s dish.

À la carte, by design

One piece of your stack.
Never the whole contract.

OpenPotluck is the serving layer. Retort is an independent agent harness. Tic is a workspace for long-running agents. Use any one, any pair, or all three.

The gateway is designed to speak the OpenAI protocol, so your choice of editor, chat interface, or agent stays yours.

  1. 01
    OpenPotluck

    The models. The hardware. The table.

    Serving
  2. 02
    Retort

    The agent that works with your models.

    Harness
  3. 03
    Tic

    The place your agents live and work.

    Workspace

A potluck for local AI

A little of yours.
A world of possibility.

We’re building OpenPotluck for people who believe
their own hardware can do more. Together.

In development · Private tables first