What is homerail?

HomeRail is a TypeScript runtime that turns quick, throwaway AI agent chats into structured, repeatable workflows you can inspect and rerun. The name reflects the idea behind it: it is meant to run on your own homelab, NAS, or home server, and work flows step by step along a graph of connected agents rather than getting lost inside one long chat conversation. The main goal is to ask as little as possible of the person using it, since attention is treated as the scarcest resource in any home automation setup. The project is built around three main ideas. Voice is the preferred way to interact with it, since speaking takes less effort than typing, though a text option is always available too. Instead of showing raw logs or technical data, the system tries to generate a simple interface for whatever is happening at the moment, though this generated interface feature is still early and evolving. Underneath both of these sits a directed graph engine that runs multiple AI agents through explicit steps, where every handoff between agents is recorded and can be replayed later, unlike a normal chat which cannot be inspected afterward. What currently works includes the graph execution engine itself, which is the most mature part, along with a command line tool called hr for starting the system, checking its health, running workflows, and reviewing results. There is also an early voice interface with speech recognition and text to speech, defaulting to Chinese, and a Docker based worker system that runs each step of a workflow inside its own isolated container. To run HomeRail you need Node.js 20 or newer, Docker, and access to a Claude Agent SDK compatible AI model. The README is written so that a coding agent like Claude Code can read it directly and follow the setup steps itself, including installing dependencies, building the project, and running a first test workflow without needing a live AI model connected yet.