Skip to main content

Your knowledge.
Your machines. Working together.

Company Brain brings selected sources, cited investigations, and human review together on infrastructure you control. That workflow is in development. Here’s how it builds on the local runtime available today and our experimental work on shared compute.

Inside Potluck · Desktop 0.1.11

Planned

From company sources to reviewed work.

Company Brain is being built for recurring work such as investigating a failed release or finding the current procedure. The intended workflow is:

  1. Sources: select relevant procedures, decisions, and prior incidents.
  2. Cited investigation: assemble a brief with evidence, hypotheses, and open questions.
  3. Human review: inspect sources, correct the brief, and record the outcome.
  4. Reuse and evaluation: propose reviewed knowledge and tests for future improvements.

This integrated workflow is in development. The runtime and experimental compute work below are its foundation. Explore the workflow.

01

Several computers run parts of the same model.

A larger model needs room for its weights: the data it uses to generate an answer. The experimental split runtime can place parts of those weights on different computers and connect their work over your private machine network. Each selected computer contributes to the same inference.

  1. Choose the main computer, called the head. It keeps the complete model file and handles your requests.
  2. Choose one to three other computers you own as workers. Enable sharing on each; they receive their assigned weights without downloading a separate model file.
  3. Check the available memory, then start the split. Potluck verifies that every selected worker received model weights before marking the model ready.

The computers need compatible split engines and must stay connected. Free memory, model size, and context length determine placement. Some weights may stay on the main computer’s CPU. Splitting adds capacity; response speed still depends on the hardware and network. If a worker disconnects, restore it and start the split again.

See the measured two-computer configuration, including exactly where the model weights ran.

02

A model runs on your computer.

Pick a model from the library and download its weights. Potluck estimates what fits your hardware and verifies the download before installing it. Your CPU or GPU then generates the answer locally.

Current stable desktop downloads require sign-in. The 0.1.12-rc.1 prerelease adds local chat, models, and memory without an account. Download a model first; local inference can then work offline. Pairing and network routes require sign-in and a connection.

03

You can also send a request to one computer.

The desktop app also supports a simpler workflow: one paired computer runs the entire model and answers a request from another. Potluck includes a WireGuard mesh: an encrypted connection between those machines, with connection setup and diagnostics built in. You do not need a separate VPN account.

  1. Install Potluck on each machine and pair them using the in-app pairing flow.
  2. On the computer that will answer, load the model and enable serving from the Network view.
  3. Check that the machine is reachable and ready, then select My Machines as the routing scope in chat.

The request can be served by an eligible paired machine with the requested model loaded. This routing workflow uses one serving machine. To spread the same model across machines, use the separate splitting workflow described above.

Follow the two-machine setup guide for the complete walkthrough.

WireGuard protects traffic between the machines. The computer generating the answer necessarily processes the prompt in plaintext. Only pair machines you control and trust.

04

Your context survives the session.

The bundled Model Context Protocol (MCP) server lets compatible coding tools save and retrieve project facts, preferences, decisions, and active tasks. Your coding agent stays in the tool you use; Potluck supplies the memory.

In Settings → Connect tools, choose a client and copy its generated configuration. Restart that client, then ask it to save a project fact. You can inspect and edit the result in Potluck’s memory browser.

Local chat selects relevant saved context. You can pin a memory to keep an applicable standing instruction in retrieval. Recognizable credential patterns are redacted, but the memory store is not a password manager.

The agent activity view displays tool-call metadata. It is an observation surface, not an agent runtime or a log of external agents’ raw prompts.

05

A wider network is the next step.

The intended routing scopes are your machines, an invited trusted circle, and an open contributor pool. The current desktop release supports local use and your own paired machines. Trusted-circle inference and a dependable open contributor network are still in development.

The coordinator code includes blinded job dispatch: account identity is left out of the job sent to a contributor. That is one part of the design, not proof that the full public service is ready.

Contribution credits, shared-network pricing, and member governance are future work. They are not required to run the local app. The roadmap tracks the remaining work.

Start with one machine.

The install guide covers supported platforms, checksums, and first-run setup.

Install Potluck ↗