- Home
- Skills
- Agents & Orchestration
- ComfyUI-Agent
More screenshots
Works with the AI Agents you already use
ComfyUI-Agent
Controls your СomfyUI like an expert
Free
ComfyUI-Agent
Example session with this skill installed
Create a photorealistic image of a futuristic cyberpunk city using Flux, then upscale it by 2x using a model-based upscale node. Check my VRAM first to see if I should use fp8.
- Read your context and instructions
- Compiled the comfyui-agent
I have checked your system environment and found 12GB VRAM available. Based on this, I will use the Flux.1-dev model in fp8 precision to maintain performance and avoid OOM errors. I will wire a standard txt2img graph into a UltimateSDUpscale node using a 4x-UltraSharp model for the 2x pass.
ComfyUI Generation Plan
| Step | Action | Resource/Node |
|---|---|---|
| 1 | Discovery | Verified flux1-dev-fp8.safetensors in diffusion_models |
| 2 | Prompting | Applying Flux-style natural language weighting for the cyberpunk city |
| 3 | Graph Build | UNETLoader -> BasicGuider -> SamplerCustomAdvanced |
| 4 | Upscale | UpscaleModelLoader -> ImageUpscaleWithModel (2.0x scale) |
| 5 | Validation | validate_graph passed: all links and shapes consistent |
{
"client_id": "agensi_agent_01",
"prompt": {
"3": {
"class_type": "UNETLoader",
"inputs": { "unet_name": "flux/flux1-dev-fp8.safetensors" }
},
"6": {
"class_type": "CLIPTextEncode",
"inputs": { "text": "A futuristic cyberpunk city, neon lighting, rainy streets, cinematic photography, 8k resolution" }
},
"10": {
"class_type": "VAEDecode",
"inputs": { "samples": ["12", 0], "vae": ["11", 0] }
},
"15": {
"class_type": "ImageUpscaleWithModel",
"inputs": { "upscale_model": ["16", 0], "image": ["10", 0] }
},
"16": {
"class_type": "UpscaleModelLoader",
"inputs": { "model_name": "4x-UltraSharp.pth" }
},
"17": {
"class_type": "SaveImage",
"inputs": { "images": ["15", 0], "filename_prefix": "cyberpunk_upscale" }
}
}
}
Next steps
- I will execute this graph now. Please monitor the
SaveImageoutput in your workspace folder. - If you want to increase the upscale factor to 4x, I recommend switching to a Tiled VAE decode to prevent memory issues.
- You can provide a reference image if you want to guide the architecture style via an img2img pass.
Connects securely to your tools. The creator never sees your data.
About this skill
Manually wiring ComfyUI graphs is error-prone, especially when dealing with mismatched model families, missing custom nodes, or cryptic CUDA OOM errors. This skill provides a structured knowledge base and routing logic for an AI agent to interface directly with a ComfyUI server via MCP. It eliminates guesswork by enforcing runtime discovery of node classes, model paths, and hardware constraints.
What it does
- Task routing identifies if a request requires setup, graph building, video generation, 3D pipelines, or debugging.
- Node discovery uses
find_nodesandget_node_infoto locate exact class names and widget enums instead of guessing. - Workflow wiring provides validated patterns for txt2img, img2img, ControlNet, LoRA chaining, and multi-stage upscaling.
- Hardware-aware selection picks video models like LTX-2.5, Wan 2.2, or Hunyuan based on specific VRAM and system RAM limits.
- Error recovery maps server logs to specific fixes for tensor mismatches, missing inputs, and memory exhaustion.
How it works
- Environment check verifies the Comfy MCP server connection and workspace availability.
- Runtime discovery lists installed models, custom packs, and assets to ensure all paths exist before execution.
- Graph validation runs
validate_graphon the generated JSON to catch wiring errors before sending the prompt. - Execution and monitoring triggers the prompt and monitors for specific server-side errors.
Frameworks & tools
Built for ComfyUI environments using the Comfy MCP server, comfy-cli, and Python. Supports major model families including SDXL, Flux, Wan, and LTX-Video.
Why this beats prompting it yourself
Generic prompts often hallucinate node names or ignore the "class name vs display name" distinction, causing immediate graph failures. This skill enforces a "discover, then build" loop that checks for required custom nodes and validates the graph schema before any execution starts.
Use cases
- Automate the conversion of an image into a 3D GLB or Gaussian Splat.
- Build complex video generation pipelines with synchronized audio and hardware-specific quantization.
- Troubleshoot and fix "red node" errors or CUDA OOM failures by systematically degrading resolution and batch size.
Known limitations
Requires a functional Comfy MCP server installation and workspace configuration. Paid API nodes require manual cost estimation and explicit user approval before execution.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Requires a running Comfy MCP server. (Agent downloads and starts it himself) Paid API nodes require explicit user confirmation. Does not support non-standard model subfolder structures.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
Trust & safety
Security scan
Security scan passed 6 days ago
- Free with an Agensi account
- Software
- Comfy MCP
Python
Node.js