ComfyUI Knowledge MCP Server

ComfyUI Knowledge

Search node schemas, browse curated API prompt templates, and validate workflow JSON offline before queueing on the user's GPU.

Workflow

1. Discover nodes — search the materialized index by name or category (sampling, loaders, image, etc.).
2. Inspect schema — fetch one node class for required/optional inputs and output types.
3. Find starters — search curated templates, then load a template by id for full API prompt JSON.
4. Validate offline — validate the draft graph before queueing on the live ComfyUI bundle to catch bad wires and unknown classes.
5. Learn patterns — read bundled guides for API prompt shape and the txt2img build loop.

API prompt shape

• Top-level keys are string node ids ("3", "4", …).
• Each node is ["NodeClass", {inputs}] or {"class_type": "NodeClass", "inputs": {...}}.
• Wires use ["source_node_id", output_index] (usually index 0).
• End graphs with SaveImage or PreviewImage so execution history and image/output fetch on the live ComfyUI bundle can retrieve results.

Gotchas

• ckpt_name and other model filenames must exist on the user's ComfyUI install — list checkpoints on the live instance before queueing.
• Custom nodes may be missing from the index — probe node schema on the live ComfyUI instance or refresh the dataset after upgrades.
• Indexed schemas come from object_info at refresh time; they differ across ComfyUI versions and installed custom nodes.