@bachstudio/comfyui-mcp-server

A dead-simple Model Context Protocol (MCP) server for ComfyUI: give it a text prompt, get an image back.

It talks to ComfyUI over HTTP, so it can point at ComfyUI on the same machine or at a ComfyUI exposed on a public URL (via a tunnel) — which is what lets a hosted/online MCP platform drive a ComfyUI running on your own PC.

Tools

| Tool | Input | Purpose | | --- | --- | --- | | generate_image | prompt (required), optional width, height, steps, seed | Text-to-image (Z-Image-Turbo). Async: returns a job_id immediately. | | edit_image | prompt + image1 (required), optional image2, image3, seed | Image editing / combining (Qwen Image Edit 2509 + Lightning 4-step LoRA). Images accepted as http(s) URL, data: URL, or raw base64. Output size follows image1. Async: returns a job_id immediately. | | generate_wan22_video | image (required), optional prompt, negative_prompt, seed, workflow_path | Video generation using the local Wan2.2 Remix V3 image-to-video ultra-smooth workflow. Images accepted as http(s) URL, data: URL, or raw base64. Async: returns a job_id immediately. | | dressup_flux2klein | model_image + garment_image (required), optional prompt, max_resolution, seed, workflow_path | 换装 (clothing swap) with the Flux 2 Klein 9B 加速版 (4-step euler, cfg 1.1). Puts the garment onto the model. Images accepted as http(s) URL, data: URL, or raw base64. Async: returns a job_id immediately. | | generate_walkway_video | front_image + side_image + back_image (required), optional seed, workflow_path | 走秀视频（快速版） — Wan2.2 RemixV2 I2V 首尾帧 · 11 秒 · 32fps · 四段硬切 workflow. Takes three full-body first frames (front / side / back of the dressed model) and produces one 11s runway walkway MP4. 相比旧的 64fps + SeedVR2 版本快得多（4090 上 30+ 分钟 → 5~15 分钟量级）。 Async: returns a job_id immediately. | | get_result | job_id (required) | Fetch the result of an async job. Returns structured progress while processing, or the public ComfyUI /view URL when done. |

Asynchronous jobs

Image generation can take long enough to exceed an MCP host's request timeout. All generation tools therefore start a background job and return a job_id right away. Poll get_result every 10-15s until it returns the image/video public URL. Jobs are kept in memory for one hour after completion and are lost if the MCP server restarts.

// 1) generate_image/edit_image -> { "job_id": "…", "status": "submitted" }
// 2) get_result { "job_id": "…" } -> { "status": "running", "stage": "...", "elapsed_seconds": 42, "progress": { "comfy_status": "running", ... } }   (repeat)
// 3) get_result { "job_id": "…" } -> "https://your-comfy.example/view?filename=...&subfolder=&type=output"

Configuration (environment variables)

| Variable | Default | Description | | --- | --- | --- | | COMFYUI_BASE_URL | http://127.0.0.1:8188 | ComfyUI HTTP API base URL (local or public tunnel URL). | | COMFYUI_AUTH | (none) | Optional raw Authorization header value, e.g. Bearer xxxxx. | | COMFYUI_TIMEOUT_MS | 60000 | Per-request timeout. | | COMFYUI_JOB_TTL_MS | 3600000 | How long completed jobs remain queryable in memory. | | COMFYUI_WAN22_WORKFLOW_PATH | bundled workflow | Path to the Wan2.2 Remix V3 ComfyUI UI workflow JSON. If unset or unreadable, the server searches the user's Desktop/Documents, then falls back to the bundled workflow. | | COMFYUI_FLUX2KLEIN_WORKFLOW_PATH | bundled workflow | Path to the 换装-Flux2Klein API JSON (ComfyUI "Save (API Format)" export). If unset, the bundled workflow is used. | | COMFYUI_WALKWAY_WORKFLOW_PATH | bundled workflow | Path to the WAN2.2 RemixV2 I2V 首尾帧 11s / 32fps API JSON for generate_walkway_video. If unset, the bundled workflow is used. | | COMFYUI_PUBLIC_BASE_URL | auto-discover | Public ComfyUI base URL used to build returned /view links. If unset, the server tries ~/.cloudflared/config.yml, then falls back to COMFYUI_BASE_URL when it is not localhost. |

Run

npx -y @bachstudio/comfyui-mcp-server

MCP client / platform config

{
  "mcpServers": {
    "comfyui": {
      "command": "npx",
      "args": ["-y", "@bachstudio/comfyui-mcp-server"],
      "env": {
        "COMFYUI_BASE_URL": "https://<your-public-comfyui-url>"
      }
    }
  }
}

Requirements

generate_image targets the Z-Image-Turbo model:

• diffusion_models/z_image_turbo_bf16.safetensors
• text_encoders/qwen_3_4b.safetensors
• vae/ae.safetensors

edit_image targets the Qwen Image Edit 2509 model with the Lightning 4-step LoRA:

• diffusion_models/qwen_image_edit_2509_fp8_e4m3fn.safetensors
• loras/Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors
• text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors
• vae/qwen_image_vae.safetensors

generate_wan22_video targets the local Wan2.2+Remix+V3 image-to-video ultra-smooth workflow. By default it uses the bundled workflow copied from:

• C:\Users\githu\Desktop\图片素材\Wan2.2+Remix+V3+图生视频超流畅最强版.json

Set COMFYUI_WAN22_WORKFLOW_PATH or pass workflow_path if the workflow moves.

dressup_flux2klein targets the Flux 2 Klein 9B 加速版 workflow. It needs:

• diffusion_models/flux-2-klein-9b.safetensors
• text_encoders/qwen_3_8b.safetensors
• vae/flux2-vae.safetensors
• Custom nodes: LayerUtility: ImageScaleByAspectRatio V2, JjkText, EmptyFlux2LatentImage, Flux2Scheduler, RandomNoise, CFGGuider, ReferenceLatent, SamplerCustomAdvanced, Int

generate_walkway_video targets the 【Work-FIsh】WAN2.2-RemixV2-I2V首尾帧视频_四段11秒界面版 workflow (32fps, 快速版，去掉 SeedVR2 超分与二次 RIFE 补帧). It needs:

• Wan2.2 high-noise / low-noise UNet weights
• text_encoders/umt5_xxl_fp16.safetensors
• Wan VAE
• Custom nodes: PathchSageAttentionKJ, ModelPatchTorchSettings, WanFirstLastFrameToVideo, ImagePanZoomSequence, ImagePanSequence, GetImageRangeFromBatch, RIFE VFI, ConcatWithBlackGap, VHS_VideoCombine, LayerUtility: ImageScaleByAspectRatio V2, ImageResizeKJv2, Int, INTConstant, easy cleanGpuUsed

Set COMFYUI_FLUX2KLEIN_WORKFLOW_PATH / COMFYUI_WALKWAY_WORKFLOW_PATH or pass workflow_path to point at a different API-format JSON.

License

MIT © bachstudio