swipi-engine

Turn a one-line prompt into a playable Phaser web game. An agentic framework on top of Claude that handles the full pipeline: classify the game → scaffold a project → write a Game Design Document → generate assets → wire config → implement code → verify it builds and runs.

Three ways to use it. Pick whichever fits your workflow.

1. Use it inside Claude Code (plugin)

Install once from inside Claude Code:

/plugin marketplace add Citronetic/swipi-engine
/plugin install swipi-engine@swipi-engine

Then, in an empty directory, drive game generation with slash commands:

Example:

/swipi-engine:swipi-new "Build a Snake clone with WASD controls and a dark theme"

When it finishes:

npm install
npm run dev        # opens at http://localhost:5173

The plugin also exposes a proactive swipi-debugger subagent that Claude delegates to automatically when a build or runtime error shows up during a session.

Full plugin reference: packages/plugin-claude-code/README.md.

2. Use it as a TypeScript library (@swipi/core)

Call the game-generation engine directly from your own code — no Claude Code, no agent runtime required.

npm install @swipi/core @anthropic-ai/sdk

import path from 'node:path';
import { AnthropicLLMClient, runClassifyScaffoldGDD } from '@swipi/core';

const result = await runClassifyScaffoldGDD(
  "Build a Snake clone with WASD controls and a dark theme",
  {
    llm: new AnthropicLLMClient({ apiKey: process.env.ANTHROPIC_API_KEY }),
    sharedDir: path.resolve('./swipi-shared'),   // templates/ + docs/
    workspaceDir: path.resolve('./games/my-snake'),
    onPhaseStart: (p) => console.log(`[${p}] start`),
    onPhaseComplete: (p) => console.log(`[${p}] done`),
  },
);

console.log(result.classification.archetype);   // "grid_logic"
console.log(result.gddPath);                     // "./games/my-snake/GAME_DESIGN.md"

Individual tools are also exposed directly for finer-grained control: classifyGame, generateGDD, generateAssets, generateTilemap, plus the Template Skill and Debug Skill evolution pipelines.

The library is provider-agnostic — ship a Claude adapter out of the box, but any LLMClient implementation works (OpenRouter, DashScope, local models via the included OpenAICompatClient, or your own).

Full library reference: packages/core/README.md.

3. Use it as a REST service (@swipi/api)

Expose the same engine over HTTP — useful for CI pipelines, web UIs, or integrating with other agent platforms.

git clone https://github.com/Citronetic/swipi-engine.git
cd swipi-engine
npm install
npm run build --workspaces
export ANTHROPIC_API_KEY=sk-ant-...
npm run start --workspace=@swipi/api
# swipi-api listening on http://0.0.0.0:3000

Endpoints

Example

# Kick off a run
curl -s -X POST http://localhost:3000/generate \
  -H 'Content-Type: application/json' \
  -d '{"prompt":"A Snake clone with WASD controls and a dark theme"}'
# → { "runId": "...", "status": "queued", "links": { ... } }

# Tail progress (Ctrl-C after you see kind:"done")
curl -N http://localhost:3000/runs/<runId>/events

# Download the generated project
curl -o game.zip http://localhost:3000/runs/<runId>/artifact.zip
unzip game.zip -d game && cd game && npm install && npm run dev

Full API reference: packages/api/README.md.

Repository layout

swipi-engine/
├── .claude-plugin/marketplace.json     Claude Code marketplace manifest
├── packages/
│   ├── shared/                         Templates and design docs
│   │   ├── templates/                    Phaser scaffolds — core + 5 archetypes
│   │   └── docs/                         GDD schema, asset/debug protocols, module manuals
│   ├── plugin-claude-code/             Claude Code plugin (skills, commands, agent)
│   ├── core/                           @swipi/core — TypeScript library
│   └── api/                            @swipi/api — Hono REST service

License

Apache-2.0. Derived from OpenGame (see NOTICE for upstream attribution).