ComfyUI Agent SKILL

ComfyUI Agent SKILL

Prerequisites: Install the CLI: pip install -U comfyui-skill-cli. The official local Comfy MCP is optional and should be connected using the official guide when MCP capabilities are needed. All shell commands must run from this project's root directory (where this SKILL.md is located).

[!IMPORTANT] Directory Sensitivity: The CLI reads config.json and data/ from the current directory. You MUST cd into the project root before running any command. Symptom: list returns [] or server status reports not found → you are in the wrong directory.

Quick Decision

• A matching registered workflow exists → use the CLI Execution Flow (Step 1-4). This is the default for repeated, batch, and latency-sensitive work.
• The request has no matching workflow, or requires live template/model/node discovery → use the official local Comfy MCP.
• The user supplies a raw workflow that needs real environment validation → use MCP validate_workflow; use CLI run --validate for a registered workflow.
• The user asks to install nodes, download models, restart, update, or switch versions → use MCP only after explaining the effect and getting confirmation.
• User says "generate image / draw a picture" → first inspect registered workflows; if none fit, discover a current MCP template instead of guessing a model name.
• User says "import workflow / add workflow" → comfyui-skill --json workflow import <path>
• User says "img2img / use this image" → first comfyui-skill --json upload <image>, then execute
• User says "inpainting / mask this area" → comfyui-skill --json upload <mask> --mask, then execute
• User says "show previous results" → comfyui-skill --json history list <id>
• User says "what failed / check job status" → comfyui-skill --json jobs list --status failed
• User says "which server has more VRAM" → comfyui-skill --json server stats --all
• User says "what nodes are available" → use MCP nodes for live ecosystem discovery; use comfyui-skill --json nodes list when only the CLI inventory is needed.
• User says "dry run / test without executing" → comfyui-skill --json run <id> --validate
• User says "open management UI" → python3 ./ui/open_ui.py

Core Concepts

• Skill ID: <server_id>/<workflow_id> (e.g., local/txt2img). If server is omitted, the default server is used.
• Schema: Each workflow has a schema.json that maps business parameter names (e.g., prompt, seed) to internal ComfyUI node fields. Never expose node IDs to the user.
• Server: One or more ComfyUI instances configured in config.json. Check health with server status.
• Two interfaces, one Skill: CLI is the efficient execution path for known workflows. Official MCP is the adaptive path for discovering and orchestrating the current ComfyUI ecosystem. They share the same business intent and must not submit the same job twice.

Official Local Comfy MCP

When the Agent host exposes the official Comfy MCP tools, call those tools directly. Start every MCP flow with server_info and use the live tool schemas instead of assuming parameters or tool availability.

When the host does not expose MCP tools but can run shell commands, use the included stdio bridge: