AI repository intelligence

FR-AI / ANALYSIS

Why it is worth attention

It bridges large language models with a professional pixel art editor (Aseprite), enabling autonomous, iterative art generation with real-time previews and undo capabilities, all running locally with no cloud dependencies beyond an OpenRouter API key.

Who it is for

Indie game developers creating pixel art assets
AI researchers experimenting with tool-using agents
Artists curious about AI-assisted pixel art workflows
Developers building LLM-driven creative tools

Use cases

Generating and refining pixel art sprites iteratively by describing them in natural language
Rapid prototyping of game assets without manual pixel-by-pixel drawing
Exploring the capabilities of vision-enabled LLMs for visual art creation
Benchmarking different models and prompt strategies for pixel art generation

Strengths

Fully local operation with only OpenRouter as external dependency, preserving control and privacy
Integrates with a real pixel art editor (Aseprite) rather than generating standalone images
Iterative loop with visual feedback (render_preview) and undo enables the model to correct mistakes
Sandboxed per-run workspaces and automatic pruning of old runs prevent disk clutter

Considerations

Requires a manually built Aseprite executable and pixel-mcp binary, making setup non-trivial
Only tested on Fedora Linux; paths and builds may differ on macOS/Windows
Vision models are needed for preview and undo; blind models produce inferior results and cannot be directly compared

README quick start

Pixel Art Lab

A local web app for making pixel art in Aseprite using LLMs.

It works by using pixel-mcp to let the model interact with Aseprite through a tool-calling loop — the model draws, renders a preview, looks at it, and iteratively fixes the issues.

OpenRouter is currently the only supported provider.

Requirements

Node 20+
Aseprite 1.3.10+, built and working
pixel-mcp, which needs Go 1.23+ to build
An OpenRouter API key

You need a working Aseprite executable before any of this does anything — pixel-mcp drives that executable, and this app drives pixel-mcp. Building it yourself from source costs nothing; the prebuilt binaries on Steam and aseprite.org are what you pay for. Either works here, as long as aseprite_path points at it.

Setup

1. pixel-mcp

Clone it into the project root and build it. That is where this app looks by default, and the directory is gitignored, so it stays out of your history:

git clone https://github.com/willibrandon/pixel-mcp.git
cd pixel-mcp && make build && cd ..

Then point it at your Aseprite executable, in ~/.config/pixel-mcp/config.json:

{
  "aseprite_path": "/absolute/path/to/aseprite",
  "timeout": 30
}

aseprite_path must be absolute — pixel-mcp does no discovery and reads no environment variable for it. Confirm the whole chain works before going further:

./pixel-mcp/bin/pixel-mcp --health

Clone it somewhere else if you prefer, and set PIXEL_MCP_BIN in .env to wherever the binary landed.

2. This app

cp .env.example .env

Set OPENROUTER_API_KEY in .env. Everything else there is optional.

npm install
npm run dev        # server on :8787, web on :5273

Open http://localhost:5273. The server listens on localhost only.

Configuration

Environment variables (all optional except the key):

Var	Default	Purpose
`OPENROUTER_API_KEY`	—	Required. Stays server-side; never sent to the browser.
`PIXEL_MCP_BIN`	`./pixel-mcp/bin/pixel-mcp`	Path to the MCP server binary
`PORT`	`8787`	Backend port
`RUNS_DIR`	`./runs`	Where run workspaces are written
`GALLERY_DIR`	`./galle

Related repositories

Similar projects matched by category, topics, and programming language.

mereyabdenbekuly-ctrl

Featured

clodex-ide

Clodex is an open-source, local-first agentic IDE that combines persistent AI tasks, code, terminal, browser, Git, models, memory, and governed execution in one Electron workspace, currently in technical preview.

AI & Machine LearningAI Agents

859

0xwilliamortiz

Featured

openclaude-improved

OpenClaude is an open-source CLI coding agent that runs on any platform and supports a wide range of LLM providers, offering the same tools and workflows as Claude Code.

AI & Machine LearningLarge Language Models

577

Jakubantalik

Featured

thinking-orbs

A React component library that renders six hand-tuned animated thought orb loading indicators on a plain 2D canvas, with two purpose-tuned sizes and automatic theme detection for AI and agent UIs.

AI & Machine LearningAI Agents

1,191