ZSeven-W/dsh-game-studio

What it does

DSH Game Studio is a plugin for DeepSeek Harness that builds small browser games from a natural-language prompt and runs them in a sidebar next to the conversation. It ships 2D top-down and 3D voxel (three.js) templates, reshapes them into the described characters, rules and screens, and rebuilds the preview on every edit.

The plugin exposes 10 agent tools for creating, building and playtesting games, generating and applying art, restoring versions, exporting a single offline HTML file, and sharing a temporary read-only link over the local network. A "My games" panel lists games in the workspace with their latest playtest screenshot and can pack them into an export wall page.

Optional AI art is generated per art slot through the OpenAI image API or the locally signed-in Codex CLI; without either, code-drawn characters are used.

Install and first game

Requirements: DeepSeek Harness 0.2.1+ (web or desktop), Node.js 24.11+, and Chrome or Edge on the same machine for headless playtests. AI art additionally needs an OpenAI key or a signed-in Codex CLI.

Install from npm into a DSH profile and start the web UI:

dsh plugin --profile web add @zseven-w/dsh-game-studio@latest
dsh web

Open a session in a workspace folder and describe a game, for example Make a small game: a kitten steals dried fish in a kitchen without bumping into the robot vacuum. The agent creates the project, builds it and shows it in the sidebar; you can play it immediately. Follow-ups like Generate 2 new pieces of art for the hero with AI; I'll pick in the sidebar. or Export it as a single HTML file. use the other tools.

Each game is a plain folder: game.json holds title, style, orientation, art slots and free-form rules passed as window.GAME.rules; game.js is the editable game code exposing window.__gameState(), optional window.__autoplay(state) and optional window.__renderSlots(size); assets/ stores per-slot art with ASSETS.json; dist/index.html is the current build.

Costs and limits

OpenAI image generation is billed per image to your key, and the plugin shows an approval card with service, image count and estimated cost before spending (confirmSpending defaults to on). Codex image runs count against the Codex plan and take several minutes each. Relevant settings such as imageProvider, openaiModel, openaiBaseUrl and confirmSpending live under config in the profile's cordis.patch.yml.

Games run in a sandboxed iframe with network access blocked, and exported files make no network requests. Prompts, and any attached reference picture, are sent to the chosen image service. The share link is read-only, local-network only and expires after a configurable time (30 minutes by default).

Playtest coverage depends on the game defining window.__autoplay for control schemes other than arrow keys. The author's own measurements are based on a 10-request sample per genre with DeepSeek-V4.1-Flash; the README states the sample is small. Development uses npm ci, npm run build and npm test, with browser-dependent tests skipped when no Chrome or Edge is found.

HotRepo's take

The repository was created on 2026-11-05 and entered HotRepo's new-project list for TypeScript on 2026-11-06, gaining 7 stars in the roughly two-hour window from 10:45 to 12:45 UTC and reaching 108 stars. There is no recorded GitHub all-language daily rank and no Hacker News discussion, so the useful signal is a very early, small-but-real uptick rather than a sustained trend.

Technically, the project is interesting because the generated output is inspectable: a folder with JSON rules, editable game code and replaceable assets, plus playtests that report errors, frame rate and screenshots. That makes it more than a prompt-to-demo toy, though its value is tied to an existing DSH setup and, for AI art, a paid image service.