Related to #11143 (ComfyUI MCP skill). This is a CLI-based alternative approach.

Related to #11143

Thanks for linking #11143 @alt-glitch — took a close look at both approaches.

The core difference is where the complexity lives.

#11143 teaches the agent to hand-craft ComfyUI workflow JSON at runtime — the agent has to assemble node graphs (["4", 0] link syntax), manage polling loops, handle urllib multipart uploads, etc. It's essentially a 346-line tutorial embedded in SKILL.md. That works for simple txt2img, but real-world ComfyUI workflows have 20-80 nodes with complex routing — no LLM reliably generates those from scratch.

This PR takes the opposite approach: the user's existing workflows are the source of truth. The CLI handles import, schema extraction, parameter injection, dependency resolution, multi-server routing, and execution history. The agent's job is reduced to:

comfyui-skill list                           # what's available?
comfyui-skill run flux-portrait --args '{…}' # execute
comfyui-skill history list flux-portrait     # what happened?

A few concrete things this handles that #11143 doesn't:

• Workflow discovery: list shows all imported workflows with param counts — the agent never sees raw node graphs
• Dependency checking: deps check tells you which custom nodes or models are missing before you hit a cryptic class_type not found
• Idempotency: --job-id prevents duplicate GPU burns on agent retries (O(1) file-existence check)
• Multi-server: remote-a100/sdxl-base routes to a specific GPU box — no env var juggling

Also worth noting: #11143 triggered the supply-chain audit 3 times due to urllib.request.urlopen POST calls in SKILL.md. This PR has zero alerts — all network I/O lives in the installed CLI package, not in the skill definition.

Beyond the CLI, this project also ships a dedicated Web UI for visual workflow management — import, parameter editing, execution monitoring, and history browsing — making it accessible to non-technical users as well.

These two approaches serve different users. #11143 is great for someone who wants to learn the ComfyUI API. This one is for shipping — the agent calls a tool, gets an image back.

Happy to add a related_skills cross-reference if both land.

Updated this PR to fully align with Hermes optional-skill conventions. The skill now includes:

• SKILL.md rewritten to match the standard structure (frontmatter, When to Use, How It Works, Core Workflow, Pitfalls, Verification Checklist)
• references/cli-reference.md — complete 27-command reference verified against the actual CLI help output
• references/api-notes.md — REST API endpoint map for debugging
• scripts/comfyui_setup.sh — workspace initialization with auto-detection of uvx vs pip-installed CLI

Noticed #17316 was opened with a similar scope. A few things worth noting:

• This PR is maintained by the comfyui-skill-cli author — the command docs here are verified against the actual codebase, not reconstructed from README/help text
• Correct attribution: author: HuangYuChuh, license: Apache-2.0 (the CLI is Apache-2.0, not MIT)
• Includes a Web UI section for visual workflow management (import, parameter editing, execution history, i18n) — not covered in feat(skills): add comfyui skill — CLI-driven image/video/audio generation #17316
• Supports both uvx (zero-install) and direct comfyui-skill (pip/pipx) invocation
• Setup script auto-detects available CLI and accepts --url for custom server addresses

Happy to iterate on any feedback from maintainers.