Agent-native image-editing MCP surface. 21 MCP tools + 2 skills (/decompose, /visual-brainstorm) for Claude Code and Cursor. v0.17.3 (tracks vulca SDK v0.17.3; 1454 tests).
pip install vulca[mcp]==0.17.3
claude plugin install vulca-org/vulca-pluginFor real image generation: either run ComfyUI locally (free) or set GOOGLE_API_KEY for Gemini. Mock mode works without either.
You: > /decompose /path/to/painting.jpg
Claude: Reads the image with view_image, authors a decomposition plan, calls layers_split(mode="orchestrated", plan=...), returns one transparent PNG per entity. Iterates per the skill's 10-branch decision tree if segmentation fails on a specific entity.
| Tool | Description |
|---|---|
generate_image |
Pure image generation (no evaluation loop) |
create_artwork |
Single-pass cultural-guided creation + evaluate |
inpaint_artwork |
Region-based inpainting |
view_image |
Read image metadata + base64 for the agent |
evaluate_artwork |
L1–L5 cultural scoring against a tradition |
list_traditions |
List all 13 cultural traditions |
get_tradition_guide |
Detailed tradition reference |
search_traditions |
Keyword search across tradition knowledge |
layers_split |
Decompose an image into semantic layers (orchestrated mode for /decompose) |
layers_list |
List layers in a session directory |
layers_edit |
Structural edits (add/remove/reorder/toggle/lock/merge/duplicate) |
layers_redraw |
Redraw one layer with new instructions |
layers_transform |
Apply transform ops to a layer |
layers_composite |
Composite layers back into a flat image |
layers_export |
Export to PNG directory with manifest |
layers_evaluate |
Per-layer L1–L5 evaluation |
brief_parse |
Parse a creative brief into structured form |
generate_concepts |
Generate N concept variation images from a prompt |
archive_session |
Archive a completed session for tradition-learning feedback |
sync_data |
Sync sessions + evolved weights |
unload_models |
Admin: release model memory (MPS/CUDA) |
| Skill | Trigger | What it does |
|---|---|---|
/decompose |
"decompose /path/img.jpg" | Loads SKILL.md, reads image, orchestrates layers_split, iterates per decision tree |
/visual-brainstorm |
"visual-brainstorm [topic]" | Turns fuzzy visual intent into a reviewable proposal.md. Zero-pixel, Discovery-metadata only. Scoped to 2D illustrative/editorial imagery. See skills/visual-brainstorm/SKILL.md. |
- Python 3.10+
uvinstalled for the default MCP runner (uvx --from vulca[mcp] vulca-mcp)pip install vulca[mcp]==0.17.3
Apache 2.0