Aseprite AI Artist

Your coding agent, painting in the Aseprite window you already have open.

Not a generated PNG. Not a file changed behind your back. The document on your screen, one pixel at a time — and every step is one Ctrl+Z.

_{☔ 500×400 · 384 frames at 24 fps · 16 s · 24 layers — drawn by Claude Opus 5.5 through this server.}

Install · How it draws · Skills · Gallery

✨ What it feels like

"Draw me a 32×32 knight in the PICO-8 palette, then give him a 4-frame idle."

You type that, and watch it happen in Aseprite: a palette, a silhouette, shading, layers, a breathing idle, a tagged cycle. The agent looks at its own work after every step, and at the end tells you what it compromised on.

You stay in charge. Don't like the helmet? Ctrl+Z, or just say so.

Works with Claude Code, omp, OpenCode, Codex CLI, Gemini CLI, Cursor, VS Code and Windsurf. Best results so far: Claude Opus 5.5.

🚀 Install

You need Aseprite 1.3+ and Node 22.6+ (macOS, Linux or Windows).

1. The Aseprite extension — then quit and reopen Aseprite:

npx @pebbly/aseprite-ai-artist install-extension

2. Your agent:

/plugin marketplace add with-pebbly/aseprite-ai-artist
/plugin install aseprite@aseprite-ai-artist

omp plugin marketplace add with-pebbly/aseprite-ai-artist
omp plugin install aseprite@aseprite-ai-artist

npx @pebbly/aseprite-ai-artist install opencode
opencode plugin add github:with-pebbly/aseprite-ai-artist#main::path:opencode

npx @pebbly/aseprite-ai-artist install codex      # or gemini, cursor, --all

3. Check it — restart your agent, then:

npx @pebbly/aseprite-ai-artist doctor

All ticks means you're ready. Updating? Run step 1 again. Odd setups (Steam, custom folders, Windows paths) are in docs/INSTALL.md.

npx @pebbly/aseprite-ai-artist serve --headless --aseprite /path/to/aseprite

🎨 How it draws

Ask /aseprite:studio for anything and it runs the whole job:

After every step that changed pixels, it looks at the result before moving on.

🔤 The grid loop

New in 0.5.0. The agent doesn't draw with circles and rectangles — it types the frame out, one character per pixel, and draws that text in one call. To fix something, it reads the canvas back in the same format and changes only the characters that are wrong.

It sees the whole shape while writing it, so silhouettes stay even, limbs keep their length between frames, and an edit never spills into its neighbours. Animation is the same trick: copy the last frame's rows, move the arm, draw.

Since 0.6.0 every character stands for one palette entry, the same on every read, so a big canvas can be read and edited in pieces without the letters changing meaning between them. A grid can be as large as the canvas.

Early signal, not a measurement: with the same model and prompts, the mage benchmark went 3 → 4/10 and the road 1 → 3/10; the tree stayed at 2 (compare them).

📚 The pixel-art knowledge base

The agent doesn't improvise a hand at 16 px or a horse's walk — it looks them up.

62 rules · ~400 pixel templates · 2006 palettes — browse them at pixeli.pebbly.space/knowledge 👀

Every rule gives size budgets, a procedure, common mistakes and grid templates the agent pastes straight onto the canvas. Any MCP client reads them as rules:// resources from rules/.

🎨 Palettes: the console classics plus Lospec's 2000 most-downloaded, each credited — preset: "endesga-32", or just a word of the name.

👀 Review by fresh eyes

New in 0.8.0. The agent that drew a sprite is the worst judge of it — it knows what it meant, so it sees it. studio hands the result to pixel-critic, which first writes a cold read (what the picture shows to someone who was never told), then a scored critique. A draft ships at 7/10 with nothing blocking; below that the agent fixes the one thing named, snapshotting first. Each fix goes back to the critic as a blind A/B — old and new under neutral names — and stays only if it wins, so a round can't quietly make good art worse.

It costs time (roughly one more critique per round) and its effect on quality is not yet measured — the thirteen experiment runs behind it are in the gallery, next to the benchmark rows.

🧰 Skills

studio picks these for you. Call one directly when you know the step you want — as /aseprite:<name> in Claude Code and omp, or just ask elsewhere.

In Claude Code and omp, four specialists take stages off the main agent: palette-smith, rig-builder, animation-director and the read-only pixel-critic.

🏆 Which model?

Live from the benchmark: every model draws the same fixed prompts and is scored against written criteria. Your model is not on it? Send us a sprite!

🗂️ Gallery and benchmark

Everything drawn with the plugin — with its prompts, models and .aseprite source — is at pixeli.pebbly.space. The benchmark puts every model and plugin version through the same three fixed prompts and scores them against written criteria, and the knowledge base shows the rulebook the agents draw by.

Made something? Ask your agent for /aseprite:submit — it packages the files and opens the pull request.

🔧 Under the hood

your agent  ──MCP──▶  server  ──▶  bridge  ──▶  Aseprite extension

• 18 tools, grouped by noun (draw, look, layer, frame, export…) instead of ninety — every tool costs context on every turn. Reference.
• A knowledge base in rules/ — 62 files from line craft to walk cycles, with pixel templates — served over MCP, so every client gets the same craft knowledge.
• Safe by default. Every action is one undo step. With Aseprite detached, tools refuse instead of editing files on disk. Everything binds to 127.0.0.1. Details · security.

🛠️ Development

pnpm install && pnpm run build
pnpm test                 # TypeScript
pnpm run test:extension   # the Lua handlers, inside a real Aseprite
pnpm gallery:check        # gallery and benchmark data
pnpm web:dev              # the site, locally

Read AGENTS.md before changing anything — it lists the rules that aren't negotiable and the Lua gotchas that already cost someone a day.

📜 Licence

MIT — see LICENSE. Aseprite is a trademark of Igara Studio S.A.; this project isn't affiliated with them.