HomeRail

HomeRail is a TypeScript runtime that turns one-off agent chats into auditable, reusable workflows, run on a homelab, NAS, or home server. Agent work flows node to node along the explicit edges of a DAG (directed acyclic graph) instead of accumulating in a single chat, so every handoff is traced and every run can be inspected, replayed, and evaluated. The design assumes a person's attention is the scarcest resource, so input is meant to be voice where possible and output a generated interface rather than logs or JSON.

The DAG runtime is the most mature part. A Manager service coordinates workflows, and a Node service provisions one Docker Worker container per DAG node, with the workers of one run sharing a workspace. Each node runs in its own context window and receives only the handoff it needs, and templates map a provider and model to each agent, so a stronger model can plan and review while cheaper models do the bulk of the work. Runs produce scorecards and evaluations, and a pattern library covers designs such as quorum, bounded ratchets, standing goal verification, and planner-worker fan-out.

There are three ways to drive it. A voice Manager Agent plans and runs a workflow from a spoken request, with Chinese as the default voice language and text always available. The hr CLI configures, runs, supervises, and replays workflows directly, and an existing coding agent such as Claude Code or Codex can operate that CLI to develop and tune DAG templates. A browser Agent UI renders the voice surface and generated widgets, which are still in exploration.

Workers use harness adapters for the Claude Agent SDK and compatible backends. For the voice Manager Agent, Codex is the recommended harness because it narrates progress from the model's reasoning stream, while an experimental DeepSeek Harness backend maps its reasoning into HomeRail thinking events. Provider credentials live in an encrypted settings store.