ComfyUI Agent Skill

This repository is an Agent Skill folder (Claude Code / Claude.ai / Agent Skills). The only required file is SKILL.md.

• 中文文档: README.zh-CN.md

Status

• Requires a local or trusted self-hosted ComfyUI server (this repo is not a hosted service).
• Not a hosted generation service; this package does not provide a ComfyUI backend.
• This package does not install ComfyUI itself.
• Stable interface: registered workflows only + CLI with structured JSON output.
• Recommended trust model: do not point this at an untrusted public ComfyUI endpoint.
• PyPI package: comfyui-agent-skill-mie (includes comfyui-skill alias).

Registered Workflows

Stable (reviewed configs in assets/workflows/*.config.json):

• z_image_turbo (text-to-image)
• klein_edit (image edit)
• qwen3_tts (text-to-speech)
• ltx_23_t2v_distill (text-to-video)
• ltx_23_i2v_distilled (image-to-video)
• ace_step_15_music (music/audio)
• qwen_image_2512_4step (text-to-image)
• anima_turbo (anime/illustration text-to-image)
• krea2_turbo (artistic/painterly realism, concept art, product visualization)
• sam3_mat_image (text-driven matting)
• liveportrait (face/expression transfer)
• qwen3_tts_clone (voice cloning)

Source of truth: the runtime registry is derived from assets/workflows/*.config.json (and the corresponding assets/workflows/*.json workflow files). If this list drifts, trust the configs and python -m comfyui generate --help output.

For Skill Users (Generate / Edit / Video / Audio)

Use this when you want to run registered ComfyUI workflows via the CLI in this repo.

• Primary entry: SKILL.md
• Workflow selection and sizing: references/workflows.md
• Full CLI contract (output paths, async submit/poll, JSON schemas, error codes): references/cli.md
• Prompt enhancement playbooks: references/prompt_enhancement/

Quick Start

Install (Recommended: pipx)

pipx install comfyui-agent-skill-mie
comfyui-agent-skill-mie check
comfyui-agent-skill-mie generate -p "a cute cat sitting on a windowsill at golden hour"
comfyui-skill check
comfyui-skill generate -p "a cute cat sitting on a windowsill at golden hour"

Alternative: uv tool install

uv tool install comfyui-agent-skill-mie
comfyui-agent-skill-mie check
comfyui-agent-skill-mie generate -p "a cute cat sitting on a windowsill at golden hour"
comfyui-skill check
comfyui-skill generate -p "a cute cat sitting on a windowsill at golden hour"

Source mode (for development / maintainers)

uv sync
uv run --no-sync python -m comfyui check
uv run --no-sync python -m comfyui generate -p "a cute cat sitting on a windowsill at golden hour"

Optional short command (after uv sync installs the project):

uv run --no-sync comfyui-skill check
uv run --no-sync comfyui-skill generate -p "a cute cat sitting on a windowsill at golden hour"

From a local wheel (for testing before PyPI publish):

pipx install dist/comfyui_agent_skill_mie-*.whl

Or install from GitHub:

pipx install "git+https://github.com/MieMieeeee/comfyui-agent-skill.git"

In tool-install mode, workflows/references are read from the installed package, while writable data goes to a per-user directory:

• Windows: %APPDATA%\\comfyui-skill
• macOS: ~/Library/Application Support/comfyui-skill
• Linux: $XDG_DATA_HOME/comfyui-skill or ~/.local/share/comfyui-skill

Short alias also available: comfyui-skill

End-to-End Example (User input → decision → workflow → result)

This project is designed for agent-first workflow selection:

• The Agent makes the primary workflow decision using SKILL.md, workflow capability metadata, and references/workflows.md.
• The CLI only provides a low-risk fallback when no explicit --workflow is chosen.

1) User input (natural language)

"帮我做一张海报，标题写「夏日促销」，风格现代简洁，主色调蓝白。"

2) Decision (what gets chosen and why)

• Task type: text-to-image
• Strong cue: poster / embedded text requirement ("海报", "标题写…")
• Prefer a stronger match over the general fallback T2I workflow
• Selected workflow: qwen_image_2512_4step
• Size choice: 704x1280 (HD portrait preset for posters)

3) CLI command (agent explicitly chooses the workflow)

In this example, the agent explicitly chooses qwen_image_2512_4step; the CLI fallback selector is only a convenience when no explicit workflow is passed.

comfyui-skill generate \
  --workflow qwen_image_2512_4step \
  --width 704 --height 1280 \
  -p "Modern minimalist marketing poster, blue and white palette, bold Chinese title text '夏日促销', clean layout, high contrast, product-focused composition"

4) Example JSON result (stdout)

{
  "success": true,
  "workflow_id": "qwen_image_2512_4step",
  "status": "completed",
  "outputs": [
    {
      "path": "C:\\\\Users\\\\<you>\\\\AppData\\\\Roaming\\\\comfyui-skill\\\\results\\\\20260504\\\\120301_<job_id>\\\\output.png",
      "filename": "output.png",
      "size_bytes": 1234567
    }
  ],
  "error": null,
  "job_id": "<prompt_id>",
  "metadata": {
    "seed": 123456789,
    "prompt_id": "<prompt_id>",
    "prompt": "Modern minimalist marketing poster ...",
    "width": 704,
    "height": 1280
  }
}

Adding Your Own Workflow (User Extension Example)

If you want to add your own ComfyUI workflow and make it available to the skill, treat it as “connecting a proven ComfyUI graph as a reviewed capability”.

1) Export a working workflow from ComfyUI

First make sure the workflow runs inside ComfyUI, then export it as API-format JSON.

2) Import into your per-user workflow registry

Run:

python -m comfyui import-workflow .\my_workflow.json --id my_custom_workflow

This writes files under your per-user skill data directory (not the packaged built-in workflow registry), for example:

<user_data_root>/custom_workflows/my_custom_workflow/
  workflow.json
  workflow.source.json
  workflow.config.template.json

3) Review and activate the config

• Open workflow.config.template.json
• Confirm the minimal required fields:
  • description, capability, output_kind, and the exposed inputs in node_mapping
• Optional: improve agent choice by adding selection metadata:
  • intent_categories, input_modes, priority, keywords_any, selection_guidance
• Rename the reviewed file to:

workflow.config.json

Only workflow.config.json is treated as an activated registered workflow. The template is not loaded.

User-added workflows are stored in the per-user skill data directory, not in the packaged built-in workflow registry, so upgrading the skill will not overwrite them.

4) Run using your custom workflow

comfyui-skill generate --workflow my_custom_workflow -p "test prompt"

If you omit --workflow, the CLI may apply low-risk fallback selection (for example, poster-like keywords) but it will not attempt to fully “understand” arbitrary user workflows. Agent-first selection remains the intended model.

Upgrade

If you installed via pipx:

pipx upgrade comfyui-agent-skill-mie

If you installed via uv tool:

uv tool upgrade comfyui-agent-skill-mie

Troubleshooting

• Run an environment doctor check (server + workflow preflight):
  • comfyui-skill doctor (recommended)
  • comfyui-agent-skill-mie doctor
  • uv run --no-sync python -m comfyui doctor (source mode)
• If the agent/skill runs inside WSL/container/sandbox while ComfyUI runs on the host OS, 127.0.0.1 may point to the runtime itself instead of the host. Try --server http://localhost:8188 or the host machine IP (and optionally persist it via save-server).
• SERVER_UNAVAILABLE: ComfyUI is not reachable at the target URL. Start ComfyUI or re-run with --server http://<ip>:8188.
• PREFLIGHT_MISSING_NODES: install/enable required custom nodes on the ComfyUI server.
• PREFLIGHT_MISSING_MODELS: download required model files on the ComfyUI server.
• NO_OUTPUT: workflow ran but no media could be retrieved; check the workflow output node and server logs/UI.
• For PREFLIGHT_MISSING_NODES, PREFLIGHT_MISSING_MODELS, or NO_OUTPUT, consult the dependency reference: references/workflow_nodes.md.

For Maintainers (Add / Review Workflows)

Maintenance docs are intentionally kept out of SKILL.md to keep the skill instructions user-focused.

• Maintainer entry: MAINTAINER.md
• Detailed workflow registration guide: references/extension.md