Download · Quickstart · How it works · Feedback
The current release is v1.8.0, for macOS, Windows, and Linux.
Type what you want to play. A coding agent builds it, and the game appears in the pane next to the conversation and keeps running while the agent works. Ask for a change and watch it land, rather than reading a summary of it.
There is no project format and no engine to learn. The agent writes plain files into a folder on your computer, using whatever it thinks fits, and anything that runs in a browser runs in the pane. Your game never has to know Hearth exists.
And the agent is whichever one you already pay for.
Every project gets one. It plays your game itself, in its own browser, and keeps notes on what it found. Before it plays again it reads those notes back. So it does not arrive blank each time. It arrives knowing where it got stuck last week, and goes to see whether you fixed it.
That memory is what lets it answer the question you actually ask after every change, and the one nothing without a history can: did that help?
Here is a real session, the second one it played on a small platformer, after the jump height went up:
It says your last change helped.
What it thought you changed. The player's jump height was changed.
Why it says that. This time I saw the score reach 1 on picture 3, while last time I never collected any coins.
Anything worse. Nothing got worse.
Still could not work out. I could not work out how to reliably land on the first platform.
Every claim is pinned to the frame it happened on, so you can go and look. Its memory is a plain markdown file in your project, so when it gets something wrong about your game you open it and fix the line. Its old sessions never get rewritten, so you can catch it changing its mind.
It is not a good player and it will not pretend to be. Anything that wants reflexes will beat it, and what it writes down is what it tried and where it failed, never a claim that it finished. It will not tell you whether your game is fun either. That part is still yours.
It plays only when you ask, and each turn runs on your own model quota. Even when you stop it mid-session, it still writes the note. More in docs/tester.md.
Claude. Hearth runs the Claude Code you are signed into, so an API key is optional. You pick the model in the composer, and switching mid-chat takes effect on your next message. There is also a per-turn effort dial for when a change deserves more or less thought.
ChatGPT. Hearth talks to the open-source Codex CLI you installed, and Codex holds the sign-in. Models come live from the binary, reasoning effort included.
Anything else. The sidebar has a terminal beside the chat: a real shell, opened in the project folder, with the hearth tools on PATH. Run whatever agent CLI you like, and the pane keeps up because Hearth watches the folder either way.
Conversations survive, too. A chat keeps the agent's working memory across app restarts and disconnects, and stopping a turn pauses the conversation rather than ending it. Live slash commands, structured questions, exact approvals, workflows and subagents come from each installed CLI. When a feature belongs to the full-screen terminal UI, Continue in CLI resumes that exact provider session without letting the native adapter run beside it. Setup, attachments, skills, and approvals are in docs/agents.md.
Every project is an ordinary folder under ~/Hearth. Move it, back it up, open
it in any editor, put it on GitHub, or walk away from Hearth entirely and take
the game with you. Nothing lives in a database or in the cloud, and there is
nothing to export because it was never trapped anywhere.
Your conversations sit in there too, so the history of how the game got made
travels with the project. Hearth's own bookkeeping stays in a .hearth folder
that is kept out of your git history automatically.
docs/projects-and-chats.md has the layout.
- docs/probe-shim.md: a game can name its entities and actions for the tester through one small file it owns. It is an offer, and a game that says nothing is a first-class game.
- docs/mcp.md and docs/cli.md: the probe is also an MCP server and a CLI for agents working outside the app.
- docs/connect-your-engine.md: games that do not run in a browser.
- docs/roadmap.md and docs/architecture.md: what's next, and how the app fits together.
- Hearth used to be a 2D game engine. That engine is preserved on the
engine-v1branch (final release v1.2.1) and no longer developed.
Dev setup, ground rules, and the AI contribution policy are in CONTRIBUTING.md. Hearth is free and open source under the MIT license.