Skip to main content

Install on one machine. Or all of them.

Run open models on a workstation you control, or route a request to an eligible paired machine using the built-in WireGuard mesh. The current app also offers local context for compatible coding tools. The integrated company investigation workflow is in development; trusted-circle inference and the open contributor pool are not yet ready.

Desktop downloads and setup

Potluck Core is free for personal and commercial use. Company operations software and supported deployments are in development; pilot scope and terms will be discussed with each organization. See pricing.

Published versions: Apple Silicon Mac 0.1.11; Linux 0.1.11; Windows 0.1.5. Intel Mac downloads are paused.

Before you start

Fig. 1 · Requirements6 items
Requirements for installing Potluck, by platform
ItemRequirement
MacmacOS 13+, Apple Silicon. Intel downloads are paused while we repair and validate the native build. Signed and notarized, so it opens normally, no right-click needed.
WindowsWindows 10+. The installer is unsigned, so SmartScreen warns once (More info, then Run anyway). We publish a SHA-256 for every build so you can verify it yourself.
LinuxA .deb for Debian, Ubuntu, and Pop!_OS, plus a portable .AppImage for other distributions. 64-bit x86 (amd64), glibc 2.35+ (Ubuntu 22.04 or newer). A compatible Vulkan driver is required for GPU acceleration.
Cross-machineThe WireGuard mesh is included, and machines pair inside the app. Enabling it may request administrator permission for a background service and network adapter. No separate VPN account is needed.
DiskModels are the bulk of it: a few GB for a starter model, more if you pick larger ones. The app shows the download size before it fetches anything, and every download is SHA-256 verified.
AccessFree for personal and commercial use. Current stable desktop downloads require an account. Version 0.1.12-rc.1 adds local use without one; connecting machines still requires sign-in. Download a model before using local inference offline.

Install

One installer per platform. The app carries everything it needs: the inference engine, the memory store, the MCP server, and the mesh. There is no step two where you install something else.

macOS

  1. Download the Apple Silicon (arm64) DMG for M1 and newer. Intel downloads are temporarily paused.
  2. Double-click the DMG, drag the app to Applications, then eject.
  3. Open it from Applications. The build is notarized, so it opens normally with no right-click. First launch unpacks the bundled engine (~10-20 seconds).

Windows

  1. Download the installer .exe.
  2. On the SmartScreen warning, click "More info" then "Run anyway." Windows builds are unsigned; the same binaries run on our own machines, and you can match the published SHA-256 before you run it.
  3. After install, Potluck appears in the Start menu.

Linux

GPU acceleration uses your installed Vulkan driver, with automatic CPU fallback when a compatible GPU is unavailable. Validated on NVIDIA RTX 2080 Ti and Intel integrated graphics; AMD hardware validation is still pending.

  1. On Debian, Ubuntu, or Pop!_OS, download the .deb and install it with sudo apt install ./potluck-ai-desktop_0.1.11_amd64.deb, then open Potluck from the app menu.
  2. On other distributions, download the .AppImage, make it executable with chmod +x, and run it directly.
  3. Turning on the mesh installs a small background service. It asks for your password once through the system polkit prompt, no terminal required.

Command line (Mac and Linux)

Prefer a terminal, or setting up a machine over SSH? The Potluck CLI runs the same local engine without the desktop app. It sets up and tests a model, manages models, connects your machines, and shows usage. It also has a terminal coding agent and can launch Aider against your local models.

Potluck CLI: Linux 0.1.10 · Mac 0.1.10

  1. brew install newtorob/potluck/potluck
    Installs the CLI, its bundled runtime, and Node.
  2. potluck setup
    Recommends a model that fits this machine, asks before downloading it, and tests a local reply.
  3. potluck login
    Starts the engine if it is not running and opens your browser to approve this terminal. Over SSH it prints a link to open on any device. With the desktop app 0.1.7 or later it uses the app's sign-in.

macOS 15 or newer on Apple Silicon, or Linux x86-64 with glibc 2.35 or newer (Ubuntu 22.04 and later). Standalone archives install without Homebrew and need Node.js 20 or newer; each has a .sha256 file beside it. Run potluck --help for every command.

First run

The app walks you through hardware, model setup, and optional machine pairing. The stable downloads listed above require sign-in. The 0.1.12-rc.1 prerelease adds local use without an account; pairing machines still requires sign-in. Model downloads need a connection; local inference can work offline once a model is installed.

  1. Hardware check. The app reads your RAM and GPU and tells you what it can run well on this machine.
  2. Pick a starter model. It downloads with a visible size and progress, hash-verified before it loads.
  3. Pair your other machines (optional). Machines on the same network find each other; a short pairing code confirms it is really yours.
  4. Readiness check. Real probes, not a green dot: can this machine actually run inference, is a VPN fighting the mesh, is the mesh port reachable.
  5. Current stable downloads require sign-in. In 0.1.12-rc.1, choose Use on this machine for local chat, models, and memory without an account; sign in to connect machines.

Bring your coding agents

A compatible MCP client, such as Codex, Cursor, Zed, Cline, Continue, or Claude Code, can use Potluck as its persistent memory: project facts, decisions, and multi-session tasks that survive across sessions. Access from another machine requires optional peer memory sharing.

You do not hand-write the config. Open Settings, then Connect tools, pick your agent, and the app generates the exact MCP entry pointing at its bundled server. Paste it where your agent expects MCP servers and restart the agent. It will see memory_search, memory_save, task_create, and the rest of the Potluck tools.

The fastest proof it works: ask the agent to remember something, then find it in Potluck's memory browser. The agent activity view shows connected MCP sessions with tool-call metadata; it does not expose external agents' raw prompt payloads.

Your machines, one network

Pairing happens inside the app. Machines on the same network discover each other; you confirm with a short pairing code. Under the hood this builds a private WireGuard mesh between machines you own, with no third-party VPN account and no config to edit.

For another machine to answer, it must be reachable, have the requested model loaded, and have serving enabled. Choose My Machines in chat to use eligible paired computers. Keep local mode selected to run inference on this computer. Trusted-circle inference and the open pool are not ready in this release.

Memory works across the mesh too, if you opt in per machine: an agent on one workstation can search memories saved on another paired machine. Peers only ever reach the narrow peer endpoints; your full memory store, tasks, conversations, and settings stay loopback-only. Peer reads are read-only; writes stay local.

Honest limits

  • The CLI shares the desktop app's engine from desktop 0.1.7. With 0.1.7 or later on the same computer, CLI commands use the app's engine and sign-in. With an older desktop app running, potluck up reports that another runtime is already listening; update the app first.
  • Windows installers are unsigned. The Mac build is notarized, so Gatekeeper opens it normally. Windows stays unsigned, so SmartScreen warns once; we publish a SHA-256 for every build so you can verify integrity without a signature.
  • Trusted-circle inference and the open pool are in development. Use local inference or My Machines in the current release. The wider contributor network is not ready for dependable public use, and the contribution toggle does not launch an open-pool contributor in this build.
  • Peer memory sharing is per-machine, all or nothing. When you opt a machine in, paired peers can search that whole store. Per-project sharing flags are on the roadmap.
  • Mesh membership is the trust boundary. Any machine paired into your mesh can call the peer endpoints you opted into. A signed-nonce peer credential replacing reachability-as-trust is planned hardening.

Where to go next

  • /how-it-works · the routing scopes, the mesh, and what a contributor can and cannot see.
  • /benchmarks · measured numbers across memory retrieval, inference, and cross-machine latency on Mac/Linux/Windows.
  • /manifesto · why Potluck is built as a member-operated network and a founder-led company.
  • /roadmap · what is shipped, in flight, designed, and to be decided.

Questions, or want to shape it?

Join the Discord