Let the Agent Play Through the Game It Just Wrote: DSH Game Studio Is Now Open Source

DeepSeek Harness has a new open-source plugin: DSH Game Studio. The author, Fini.Yang, says on X that it is based on DSH creation mode and also incorporates his understanding of Kimi's game approach. The goal is to make a playable small game from a single sentence. The project URL and installation commands are at the end of the article.

The workflow is not complicated. You describe the game in plain language in the chat, the Agent builds a version from a template, and the right sidebar is immediately playable. There are two built-in templates: 2D top-down and 3D voxel world, with the latter using Three.js. Characters, rules, and levels can all be changed as you chat.

Auto-playtesting is worth a look. After writing the code, the Agent launches the latest version in headless Chrome or Edge and reports the stage, score, frame rate, script errors, and screenshots. Games controlled with the arrow keys are run directly; games such as grid-based, card, aiming, and rhythm games need to define window.__autoplay before playtesting can enter a real loop.

Art does not burn money from the start. The template first draws characters in code, enough to start playing. When you want to swap in images, the Agent uses the OpenAI image API or the local Codex CLI to generate 1 to 4 candidate images for each art slot, and images attached in the chat can be used as references. Before generation, an approval card appears stating the service, number of images, and estimated cost; money is spent only after confirmation. In the sidebar, you replace by number, and if you are not satisfied, you can revert to the previous version.

The rest is engineering detail: each change saves a version and can be rolled back as a whole; a read-only link is opened on the LAN, a QR code appears in the sidebar, and you can scan it with your phone to playtest; it closes after 30 minutes by default. Finally, export a single HTML file with no network requests, and it runs as soon as you drop it on static hosting. My games in the sidebar can directly open games in the workspace without going through the Agent and without consuming tokens.

The README provides a small sample test: the environment is DSH plus DeepSeek-V4.1-Flash, each game starts from one sentence and is playtested and scored independently. Collection/dodge games are the template's home turf: all 10 requests were playable on the first try, with a median of about 14 minutes. Other types include match-3, Sokoban, tower defense, basketball, platformer, rhythm game, racing, card game, farming, and coffee shop management; all 10 were also playable, with a median of about 26 minutes. The sample is small, so do not treat it as a universal promise.

The easiest thing to overestimate in AI game making is generating code; the easiest thing to underestimate is whether it is playable after generation. Many AI game demos stop at screenshots and screen recordings, and real delivery gets stuck on errors, frame rate, and mobile. DSH Game Studio puts playtesting, version rollback, and export into one workflow, and this direction is more substantial than adding another template.

The materials also include the author's earlier recommendation of Tripo. He mentions that Kimi's official game approach uses Tripo, and DeepSeek does not provide third-party 3D modeling capabilities. This background explains why he cares about 3D assets, but do not mistake the invite link for a DSH Game Studio feature. The workshop first draws characters in code, then calls OpenAI or Codex to generate images; professional 3D modeling is a separate track.

To install, you need DSH 0.2.1 or later, Node.js 24.11 or later, and Chrome or Edge on your machine for playtesting. AI art requires either an OpenAI API key or a login to Codex CLI. You can still play without it; the characters are just drawn in code first.

Installation commands:

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

Project URL: https://github.com/ZSeven-W/dsh-game-studio

If you want to try it, first have it make a collection/dodge small game, then look at the playtest report. On the template's home turf, the odds of failure are a bit lower. Once that runs smoothly, try non-template types such as match-3 and tower defense; the median time of 26 minutes is roughly what to expect.