sprite-canon

MCP server that keeps AI-generated game sprites looking like ONE game.

One character, three outfits — the blue and red rows are sprite_repaint calls, not regenerations. Same shading order, same silhouette, same result every time.

AI generators are great at making a pretty sprite and terrible at making it match the last one. Ask for the same character twice and the palette drifts, the outfit mutates, the new hat floats 3 pixels above the head — each asset is fine alone, and the game looks wrong assembled. Regenerating "until it matches" doesn't converge; it burns money and you can't diff the result.

sprite-canon takes the opposite approach, extracted from a real game project that generated ~4,000 frames and learned every lesson the hard way:

1. Your consistency rules become data — a sprite-canon.json ("the canon") holding the palette, named colour regions (skin, outfit, outline…), relative scale, and check thresholds. Committed next to your assets.

2. Verification is numeric, not visual. You cannot eyeball 96 outfit variants × 8 directions × 4 frames. sprite_verify returns hard pass/fail numbers for the defects that actually ship: off-palette pixels, accessories that jitter between frames, a region that's bright from behind and dark from the front, a repaint that touched the face.

3. Fixes are deterministic pixel operations, not regeneration. Repainting a region onto a new colour ramp preserves shading and silhouettes, never touches protected regions, and produces the same output every time. An outfit variant is one tool call, not a prompt lottery.

Install

Claude Desktop — one file, no config

1. Download sprite-canon.mcpb from the latest release.

2. In Claude Desktop, open Settings → Extensions (☰ menu → File → Settings on Windows).

3. Drag the .mcpb file into the Extensions page, review, and click Install.

(Double-clicking the file also works if your OS has the .mcpb association registered — drag-and-drop always works. Alternative: Extensions → Advanced settings → Install Extension → pick the file.)

That's the whole install: the bundle ships its own dependencies, and Claude Desktop provides the Node runtime. Requires the Claude Desktop app — for Claude Code see below.

Claude Code / other MCP clients

git clone https://github.com/useka12-eng/sprite-canon
cd sprite-canon && npm install

Then register in your project's .mcp.json (or any MCP client config):

{
  "mcpServers": {
    "sprite-canon": {
      "command": "node",
      "args": ["/path/to/sprite-canon/src/mcp/server.mjs"]
    }
  }
}

Requires Node 18+. No native dependencies — the PNG/GIF codecs are self-contained.

Build the bundle yourself

npx @anthropic-ai/mcpb pack . dist/sprite-canon.mcpb

Related MCP server: mcp-spritesheet-forge

Tools

Inputs can be PNGs, animated GIFs, or PNG spritesheets (cellW/cellH).

The workflow

canon_init      → learn the palette from your existing good assets
canon_learn     → sample skin / outfit / outline once; mark face + outline protected
sprite_measure  → read the numbers before placing anything ("where do the eyes start?")
sprite_repaint  → make variants deterministically (outfits, teams, seasons)
sprite_verify   → prove it: face untouched, nothing left over, no jitter, on palette
sprite_sheet    → look at the result as a sheet, zoomed, before it enters the game

Lessons this tool encodes

These are not hypothetical — each one shipped as a real defect first:

• Measure, don't assume proportions. A hat brim placed at "52% of head height" landed exactly on the eyes: on a 20px head the eyes are 7–9px from the top, so every fixed ratio hits them. sprite_measure reports where the face actually starts, per frame.

• Repaint with a fixed luminance range. Normalising per image maps the same source colour to different outputs depending on how much of the region is visible — our hat was bright from behind, dark from the front. The canon records each region's range once; repaint always uses it.

• Protect regions structurally. "Be careful around the face" fails at scale. protected: true means repaint cannot touch it and verify proves it didn't.

• Patch GIF palettes, don't re-encode. An indexed GIF's colours live in its colour tables — global and per-frame local ones (patching only the global table is the classic half-fix). Substituting table entries re-dresses every frame in perfect sync with zero loss.

• Region definitions have gaps; census them. 12 stray pixels of the old colour surviving a repaint is invisible to the eye and obvious to leftover. When it fires, canon_info's census shows which colours your regions don't cover.

The scale table

sprite_verify's scale check reads canon.scale.heights — relative sizes in units of a reference asset (the entry equal to 1). No tool writes this section yet; add it to sprite-canon.json by hand:

"scale": { "heights": { "hero": 1, "house": 3.4, "chicken": 0.45 } }

Then verify with scaleNames mapping file basenames to those keys. This catches the classic "the house is smaller than the hero" a week before your players do.

Practical notes

• Always pass canonPath (or a file the canon sits above). A stdio MCP server's working directory belongs to the client, not your project, so the tools refuse to guess from cwd.

• Codec limits: PNG must be 8-bit, non-interlaced, RGB/RGBA/palette (the common pixel-art cases; 16-bit or interlaced files are rejected with a clear error). The GIF encoder is exact up to 255 opaque colours per file — beyond that, nearest-palette snapping.

• sprite_sheet returns the image inline up to ~800 KB; larger sheets return the file path only.

• Spritesheets round-trip cell-for-cell: empty cells stay empty, nothing is compacted.

What this is not

• Not a generator. Pair it with whatever makes your art (PixelLab, Aseprite, Gemini, hand pixels); sprite-canon is the layer that keeps the results coherent.

• Not an atlas packer / collision tool — sprite-tools covers that well.

• Not magic: you spend ~10 minutes once per project teaching it your canon. That investment is exactly what makes every later check and fix trustworthy.

Development

npm test          # unit + end-to-end MCP tests (22)

The test suite includes regression tests for every bug an adversarial multi-agent review found in v0.1 — sheet cell compaction, GIF disposal semantics, fake-success responses, silent zero-check passes. If one fails, a bug that already existed once is back.

MIT