Skip to content
← All projects

Loopforge

Early — not production ready

An independent, local game-development agent with a desktop workbench.

Loopforge works inside a normal game repository: it turns ideas into playable experiments, gathers technical and human evidence, and helps make explicit keep, kill or refactor decisions. The Agent is the control plane and the desktop Workbench is the interface; the CLI and Skills are internal capabilities rather than things you drive by hand.

Overview

Opening a project is the whole setup: the Workbench takes a game repository through the native folder picker, then starts or reconnects that project’s Agent and loads a constrained slice of project context — without exposing provider credentials, environment variables or access tokens.

The interface is organised around the project rather than around the Agent process. A project menu switches repositories and project-level views, the header carries project identity and its actions, the work area hosts mode-specific tools, and a floating toolbar moves between exploration, design, build and test. The chat sits beside the work instead of replacing it, because project state and evidence live in the game repository — chat history is not the only record of what happened.

The boundaries are deliberate. `apps/agent` holds the domain Agent and everything a user sees; `cli` holds deterministic project operations and a headless adapter, explicitly not the control plane; `skills` holds versioned capabilities for work that needs judgement; `contracts` holds the versioned schemas. Kura supplies the generic model, session and runtime behaviour and carries no Loopforge state, and the release application embeds the Agent with a pinned Kura sidecar.

What it does

  • A desktop Workbench (Tauri + React) as the product surface, with the Agent as its only sidecar.
  • Projects open from a normal game repository; credentials and environment variables stay out of the project context that is loaded.
  • Mode toolbar for exploration, design, build and test, with the Agent chat beside the work rather than in place of it.
  • State and evidence are written into the repository, so the record of what happened does not depend on a chat log.
  • Deterministic operations and versioned Skills remain available for automation and diagnosis — internal capabilities, not the way in.
  • Generic model, session and runtime behaviour comes from Kura as a pinned sidecar, keeping domain logic on the Loopforge side of the line.

Get started

Requirements: Git, Node.js 22 and pnpm, Rust and Cargo, Python 3.11+ with uv; Godot 4 only for the Godot workflow. There is no packaged release yet, so the desktop app is built from a checkout.

Run the Workbench from a checkout

shell
$ git clone https://github.com/dopejs/loopforge.git && cd loopforge && ./dev.sh