PotluckAI

Local-first AI on a member-operated mesh. Your machines first, your circle next, the open pool when you want it.

[PotluckAI @PotluckAI]

Potluck is live.

Run open AI models on hardware you already own. Pair your machines into one private mesh and run models bigger than any single one. Your machines, a trusted circle, or the open pool.

Mac, Windows, Linux. Free → https://t.co/V3gofVIZ52

[PotluckAI @PotluckAI]

Run the model on your desktop. Chat from your laptop.

Potluck connects your computers over an encrypted mesh, putting the hardware you already own to work.

Available now for Mac, Linux, and Windows.
https://t.co/V3gofVIZ52

[Rob Newton @strakedev]

We already have the computers.

Imagine what we could build together.
We're building Potluck to connect machines in our homes, offices, and communities to power larger local AI models.

So bring what you've got. And let's build something bigger.

https://t.co/9ycqrOxLSt

[PotluckAI @PotluckAI]

Your laptop. Your desktop. Your home server.

We’ve refreshed the Potluck site to show how they fit together, what you can run today, and what we’re building next.

Take a look → https://t.co/V3gofVIZ52

[Rob Newton @strakedev]

Now we're cooking with oil. Powered by @PotluckAI !!

[Rob Newton @strakedev]

We ran a 32B coding model across a 16 GB Mac and a Linux PC with an 11 GB RTX 2080 Ti.

This was our lab benchmark. We turned that result into something people can use reliably. This means solving networking, memory placement, scheduling, and trust together. Here’s how we’re approaching that infrastructure @PotluckAI

[Rob Newton @strakedev]

Hey, cool I have replies working now and @PotluckAI is sending trade agreements.

[Rob Newton @strakedev]

And now, I'm working on a game that uses Potluck to control the AI of other nations. Running locally on your machine or you can use our cloud instance. Not released yet but here's a preview.

[PotluckAI @PotluckAI]

We ran a 32B coder model that fits on neither of our machines (a 16 GB Mac, an 11 GB 2080 Ti) across both of them.

107 tok/s prompt, 6.7 to 10 tok/s decode, over Wi-Fi. Lab numbers, not an app feature yet. This is what Phase G of Potluck is built on.

[PotluckAI @PotluckAI]

Potluck 0.1.4, Mac and Linux, auto-updating. We ran an audit on our own code and shipped four fixes: circle chats stay in the circle, history off writes nothing, failed replies show as failed, CLI file tools stay in your workspace. Notes in Discord. Windows next.

[PotluckAI @PotluckAI]

Adding new features to the coding CLI. An actual view that you can check, manipulate, update tasks, etc. All from the Potluck UI

[PotluckAI @PotluckAI]

The team is hard at work building some new features for the cli coding features. Exciting things coming!!

[PotluckAI @PotluckAI]

Rule we set before writing a line of code: nothing else gets installed.

No Docker. No Tailscale. No Python on your machine. No terminal. No account at some other company to make the networking work. One installer, WireGuard built into the app.

Adding your second machine is a toggle.

https://t.co/tL6qsg3xs0

[PotluckAI @PotluckAI]

Genuine question for the local model crowd. What's the biggest model you run that's actually fast enough to sit through?

Our 2080 Ti does Llama 3.1 8B at 43 tok/s. Same card on a 70B, most layers spilled to CPU, 2.4 tok/s. Technically running. Nobody's waiting around for that.

Where's your ceiling?

[Rob Newton @strakedev]

@PotluckAI is an official @OmarchyLinux plugin now! https://t.co/lrbVwzsvAP

The plugin process is fantastic @dhh , so far i'm enjoying Omarchy a ton. I can't wait to play with it more while on paternity leave.

[PotluckAI @PotluckAI]

People keep asking where the vector database is.

There isn't one. SQLite FTS5 keyword search fused with bge-small embeddings, RRF at k=60. If the embedder falls over it drops to keyword-only instead of erroring.

p50 10.6ms, p95 35ms on an M2 Pro 16GB.
Method and full results: https://t.co/iZZ7xEPJkh

[PotluckAI @PotluckAI]

Potluck ships an MCP server that gives coding agents memory across sessions.

Claude Code, Cursor, Codex, etc. It is one process it never calls an LLM and never reaches a vendor cloud.

The Agents view shows every call as it lands. Tool name, scope, result count, latency. Arguments are filtered to a metadata whitelist, so a query string never gets recorded.

https://t.co/tL6qsg3xs0

[PotluckAI @PotluckAI]

Potluck 0.1.3 is out for macOS, Windows, and Linux.

It runs open-weight models on machines you own. Every request routes to one of three scopes: your own machines, a trusted circle, or the open pool.

New in 0.1.3: a Usage view that reports what you consumed and what you served for others, computed on-device. Metadata only, nothing uploaded.

Download: https://t.co/tL6qsg3xs0
Benchmarks: https://t.co/iZZ7xEPJkh

[PotluckAI @PotluckAI]

We tried Llama 3.1 70B on the strongest box in our bench fleet. 13700K, RTX 2080 Ti, 62 GB of RAM.

It put 20 of 80 layers on the GPU, the rest on CPU, capped context at 2,048 tokens, and produced 2.4 tok/s.

Technically possible, unusable for chat. The gap to a usable 70B is 10 to 20x, and no single consumer card closes it.

https://t.co/iZZ7xEPJkh

[PotluckAI @PotluckAI]

The useful thing today isn’t a token or a grand decentralized-AI promise. It’s this:

Install Potluck on your laptop and gaming PC. Pair them. Ask from the laptop. Run the model on the machine with the GPU.
No cloud API. No Tailscale. No prompt leaving your machines.

Mac, Windows, Linux. Free:
https://t.co/aBW4xoZXeU

[PotluckAI @PotluckAI]

@MichaelDell This is incredibly awesome, great work!