Repos & Dungeons: watch your agents fight their way through your codebase

preview.redd.it/350kxbf280uh1.png I got bored working on a side project recently and was thinking about how old-school, pre-AI coding used to feel like playing a game (well, kinda). Tracking down a mysterious bug and finally getting the tests to pass gave the same dopamine hit as clearing a dungeon or finishing a side quest. I still see coding as a bit of a game, so I built this with Claude Code. It's absolutely useless for real work, but it's fun to watch AI agents physically fight their way through a codebase. What it is Watching Claude Code usually just means staring at terminal text. Repos & Dungeons is a free, local, open-source viewer that turns your repo into a pixel-art dungeon and your Claude Code sessions into a live RPG party. The map: folders are rooms, files are tiles. It's deterministic, so your repo always looks the same. Fog of war: unread code stays dark, so you can see exactly what the agent hasn't looked at. Combat: failing tests spawn slimes in the failing file's room. Passing tests kill them. The party: each model is a different class (Opus is a knight, Sonnet a squire, etc.), and subagents join as party members. Roleplay: they talk in D&D speak, written live by Claude Haiku ("A bug in token.ts betrays us!"). Mechanics: your context window is a torch meter. Compacting brings the fog back. Timelapse: one click replays the whole session as a 30-second MP4 or GIF. How I built it with Claude Code Design: I started with a brainstorming session in Claude Code. It asked me questions, mocked up three art styles in a browser so I could pick one (I went with the dark pixel style and asked for Claude-themed classes), then wrote a full spec. Building: Claude Code split the work into plans (parser, map generator, renderer, sound, timelapse) and built each one test-first. It ended up with about 170 tests. Reviews: after each stage, a separate Claude review agent read the whole change and caught real bugs. For example, a malformed URL could crash the local server, and grep output could leak lines of code into the game events. Claude fixed each one with a test. Testing the look: Claude ran the app in a real browser (Playwright), took screenshots, and fixed what looked wrong. It also generated the pixel art, music, logo and demo GIFs in code. At runtime: the speech bubbles come from Claude Haiku via your own Claude Code login ( claude -p ). That's roughly one short call per hero every 15 seconds; --no-narrator turns it off. My part was the idea, the design calls, and a lot of "no, make it look more like a real game." Free to try (Node 20+, run it in any repo where you use Claude Code): npx github:dbl8005/repos-and-dungeons Or add --demo to watch a preview without an agent. The first run takes about a minute to build, then it's instant. It handles 100k-file monorepos at 60 fps and runs entirely locally: the server only listens on localhost, and it reads your file tree and Claude transcripts, never your code. Repo (MIT): github.com/dbl8005/repos-and-dungeons