What we are standing on

Three of the twelve parts of Lumbridge began as someone else’s code, and all twelve depend on work we did not do. This post says whose.

The ground: Ubuntu, Rust, GPUI and Ratatui

Lumbridge is not a Linux distribution. It installs on a standard Ubuntu long-term-support release and leaves the system underneath alone. We can do that because Ubuntu gives us the same base on a laptop, a build server and a GPU box. Our own description is “bare metal, then Ubuntu, then Lumbridge”, and the middle word is not ours.

Most of the stack is written in Rust. Core, Notes, Mail, the Library, Computer and Speak are Rust programs. Notes and Computer each ship as a single binary. The language and its crates are a large part of why a small lab can maintain twelve programs at all.

The App and Mail are desktop programs drawn with GPUI, the UI framework from the makers of the Zed editor. It draws with the GPU, and it is why those two windows are Rust all the way down.

Notes and the TUI run inside a terminal and are drawn with Ratatui. In Notes, rendering is a pure function from application state to a Ratatui frame. Both programs work over SSH, because a terminal is all they need.

Herdr

Core, the App and the TUI began as forks of Herdr, an Apache-2.0 terminal runtime for coding agents with a desktop app, a mobile app and a TUI built on it. The central idea in our runtime is theirs: a background server keeps terminals alive, and clients come and go.

We now maintain our versions separately. They have changed to fit the rest of our system, and the mistakes in them are ours. Herdr is released, runs on more platforms than we test, and is ahead of us on the terminal itself. If a runtime for coding agents is all you need, look at Herdr first.

Our three repositories stay on Apache-2.0, the same license as upstream, and ship the upstream notices.

The model our agents run on

Our own agents run on an open-weight model from a large chip maker, in its official 2-bit build. We did not train it. It runs on a GPU box we own, started and supervised by Lumbridge Computer.

Open weights are what make “your machines, your data” possible for models as well as files. The same goes for the open speech models behind Voice and Speak, and for vLLM, the engine that serves our transcription model.

What we learned from Omarchy and Orca

Omarchy is an opinionated Linux distribution built on Arch and Hyprland. It showed that a set of choices on top of Linux can be a product people love, and not only a folder of config files. Lumbridge arrives the same way: one layout, one set of keys, sensible defaults, yours to change. Omarchy does this for a developer’s machine. We are trying it for a company. We support what they are doing.

Orca is an MIT-licensed desktop app that runs coding agents in parallel, each in its own git worktree, with diff review for the person in charge. Isolation per agent and review by a person are both ideas we took seriously. Our Factory prepares a git worktree per ticket, and a person merges.

Smaller debts

• tree-sitter cuts code into whole functions and types for the Library.
• SQLite holds the Mail store and the Factory’s tickets. Postgres is the Library’s index.
• Tauri wraps the phone app for Android and iOS.
• SSH carries Core’s protocol between machines.
• ffmpeg encodes audio for Speak, and curl does the HTTP for Notes.

This list is short on purpose and so it is incomplete. The notices that ship with each repository are the full record. If we have used your work and credited it badly, tell us on X or through the contact form at /apply, and we will fix it.