MCP server that lets an AI assistant draw in Aseprite over a local WebSocket bridge
Server has 9 tools with complete input schemas and descriptions. Naming follows verb_noun convention (create_sprite, draw_pixel, fill_area, add_layer, export_png, draw_pixels_batch, draw_line, draw_rect, draw_ellipse). Descriptions are present and substantive (avg ~150 chars), exceeding the 10-char minimum. However, output schemas are not documented, LLMs cannot infer what these tools return. Parameters lack format constraints (e.g., color hex validation, coordinate bounds). No error handling guidance visible. Missing tool annotations (destructiveHint, idempotentHint). All tools are WRITE operations but lack explicit state-change warnings in descriptions.
Add a new layer to the active sprite.
Create a new sprite in Aseprite.
Draw an ellipse inscribed in a BOUNDING BOX on the active layer, outlined or filled. Takes the same coordinates as draw_rect.
Draw a 1px line on the active layer — one call instead of a draw_pixels_batch triple per pixel. The coordinates are the two ENDPOINTS, both inclusive; identical endpoints draw a single pixel. Uses Aseprite's own line tool, so the stepping matches drawing interactively. A line running past the canvas edge is clipped.
Set the color of a single pixel in the active sprite.
Draw many pixels in one call — use this instead of repeated draw_pixel calls. Encode pixels as a semicolon-separated string: 'x,y,RRGGBB;x,y,RRGGBB;...'. Example: '0,0,FF0000;1,0,FF0000;2,0,00FF00'. No # prefix on colors.
No output schemas documented. LLMs cannot infer return types or plan downstream operations. Tools return success/error but structure is undocumented.
Color parameter lacks format validation. Description says 'Hex color string, e.g. #FF0000' but draw_pixels_batch explicitly forbids # prefix ('no # prefix on colors'). Inconsistent guidance invites LLM errors.
No coordinate bounds documented. Parameters x, y accept integers but min/max are not specified. LLMs may pass negative or out-of-bounds coordinates without guidance.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 67 | <=2025-11-25 | v2 |
Draw a rectangle on the active layer, outlined or filled. The coordinates are a BOUNDING BOX, not two corners: x,y is the top-left pixel and width/height are sizes, so x=2, y=2, width=8, height=6 covers columns 2-9 and rows 2-7. Set fill true for a solid rectangle, false or omitted for a 1px outline. Width and height must be at least 1. Parts of the rectangle past the canvas edge are clipped.
Export ONE frame of the active sprite, composited from its visible layers, to a PNG file. Writes exactly one file at the path given, whatever the sprite's frame count. The frame defaults to the active one; pass frame to export another without moving the active frame first. Frame numbers are 1-based — list_frames reports the valid range. To get every frame in one image instead, use export_spritesheet.
Flood-fill an area in the active sprite starting from (x, y).
All tools are destructive (WRITE) but lack explicit state-change warnings in descriptions. LLMs need to know which calls are irreversible and require confirmation.
No tool annotations (destructiveHint, idempotentHint, readOnlyHint). MCP 2.x spec supports these; their absence forces LLMs to infer intent from descriptions alone.