ComfyUI Studio

🧪 Beta test build — v0.2.0-beta.1. The official PR (MiniMax-AI/MiniMax-Code-Plugins#15) is open and waiting for review. To participate in the beta: install the standalone single-plugin repo (antianqi/comfyui-studio) following the instructions in BETA.md, then report issues back so we can fix them before the official merge.

A generic, dependency-free toolkit for driving a local ComfyUI server. Two parts in one Plugin: A. a natural-language control layer for submitting any ComfyUI workflow, polling the queue, and downloading outputs, and B. six preset workflow templates that cover the most common recurring tasks (selfie, mimicry, image edit, image fusion, drama first frame, drama image-to-video). Bundles four Skills and a stdio MCP server. Ships zero native binaries, zero installers, zero third-party dependencies. Apache-2.0.

The two parts

Part A is the workhorse. Part B is six production-ready starting points built on top of Part A.

Part B — the 6 preset scenarios

Numbered by how a user typically says it. The number is the trigger scenario, not a model id — the agent reads the user's intent and picks the matching number.

Routing rule of thumb: scenarios 1–4 → comfyui-character; scenarios 5–6 → comfyui-drama; anything else ("use ComfyUI to ...", "submit this workflow", "check the queue") → comfyui-workflow (Part A).

The detailed recipes (model stack, custom-node requirements, prompt structure, knobs) live in each Skill's SKILL.md. The table here is the index.

Part A — natural-language control

No matter what the user wants, the last three steps are always the same:

1. Submit a workflow → POST /prompt, capture prompt_id
2. Poll the queue → GET /history/<prompt_id> until terminal status
3. Download outputs → GET /view?filename=...&subfolder=...&type=output

Two equivalent entry points. Pick whichever the host agent supports.

A1. Python CLI (no MCP required)

# Health probe
python skills/comfyui-workflow/scripts/submit_workflow.py --probe

# Submit any workflow (one of the bundled presets or the user's own)
python skills/comfyui-workflow/scripts/submit_workflow.py \
  --workflow plugins/antianqi/comfyui-studio/workflows/flux2-klein-image-edit.json \
  --prompt "the same person sitting in an office chair, professional lighting" \
  --filename reference_face.png \
  --output-dir ./out

The script understands three marker conventions so you don't have to rewrite the workflow JSON for every run:

A2. stdio MCP server (no Python required)

mcp.json registers a 200-line stdio JSON-RPC server in plain Node stdlib. Zero npm dependencies. The three tools mirror the Python script:

If the host agent is MCP-aware, prefer this. If not, use the Python CLI — they are behaviourally identical.

Model boundary (read this first)

The Plugin draws a clean line between public generation models and user-supplied LoRAs.

The exact filenames / placeholders in the six bundled workflows:

If your ComfyUI install uses a different filename for any of these — for example, sd_xl_base_1.0.safetensors instead of Z-Image-Base-8steps-豹豹喵呜の白玉v2White_Marble-AIO_v2-bf16.safetensors, or ltx-2.3-distilled.safetensors instead of 10Eros_v1-fp8mixed_learned.safetensors — edit the relevant ckpt_name / clip_name1 / clip_name2 / vae_name / control_net_name field in the workflow JSON. ComfyUI will accept whatever file is on disk under the name you set.

Try it

After installing the Plugin into MiniMax Code:

Use comfyui-studio to verify that my local ComfyUI is reachable, then submit scenario 3
(改图 / flux2-klein-image-edit) with the prompt "the same person sitting in an office
chair, the same outfit, professional lighting" and report the saved image path.

Expected result: the agent probes ComfyUI, submits the bundled workflows/flux2-klein-image-edit.json, polls until done, downloads the image, and tells you where it was saved. The end-to-end walkthrough is in examples/minimal-run.md.

What you get

comfyui-studio/
├── plugin.json                       # Plugin manifest (required)
├── mcp.json                          # Registers the stdio MCP server
├── server.mjs                        # 200-line zero-dep stdio JSON-RPC server
├── LICENSE                           # Apache-2.0
├── README.md                         # This file
├── skills/
│   ├── comfyui-studio/               # Routing layer: numbers 1-6 + freeform → sibling Skill
│   ├── comfyui-workflow/             # The basic transport (Part A)
│   │   ├── SKILL.md
│   │   ├── references/api-reference.md
│   │   └── scripts/submit_workflow.py
│   ├── comfyui-character/            # Scenarios 1, 2, 3, 4 (selfie + mimicry + Klein edits)
│   │   ├── SKILL.md
│   │   └── references/
│   │       ├── prompt-patterns.md
│   │       └── lora-guide.md
│   └── comfyui-drama/                # Scenarios 5, 6 (first-frame + image-to-video)
│       └── SKILL.md
├── workflows/                        # The 6 preset workflow JSONs
│   ├── selfie-text-to-image.json     # scenario 1: 生图
│   ├── selfie-mimicry.json           # scenario 2: 模仿
│   ├── flux2-klein-image-edit.json   # scenario 3: 改图
│   ├── flux2-klein-image-edit-dual.json  # scenario 4: 融合
│   ├── drama-first-frame.json        # scenario 5: 首帧
│   └── drama-image-to-video.json     # scenario 6: 出片
├── examples/
│   ├── README.md
│   └── minimal-run.md                # 5-minute end-to-end walkthrough
└── docs/
    ├── security-notes.md
    └── troubleshooting.md

Four Skills, one routing table

The Plugin's comfyui-studio Skill is the entry point. It reads the user's intent and routes to one of three siblings. Each sibling's SKILL.md is the actual implementation guide.

Read skills/comfyui-studio/SKILL.md for the full routing table.

Requirements

• A running ComfyUI server reachable at COMFYUI_URL (default http://127.0.0.1:8188).
• The checkpoint / VAE / CLIP / ControlNet files referenced in each preset (see the model boundary table above). Edit any loader field that does not match your install.
• For scenarios 1 and 2 (comfyui-character selfie workflows): a face LoRA and a style LoRA in models/loras/. The Plugin does not bundle them; it uses placeholder names.
• For scenario 2 (selfie-mimicry.json): the custom-node pack list in skills/comfyui-character/SKILL.md (controlnet_aux, LLaMA-CPP, ModelPatchLoader, etc.).
• For scenarios 3 and 4 (comfyui-character Klein workflows): the Flux.2 Klein 9B UNET, the Qwen3 8B CLIP, and the bundled VAE.
• For scenarios 5 and 6 (comfyui-drama): two character LoRAs (for two-character dramas), an LTX-Video checkpoint with its DualCLIP text encoders and VAE, a TTS pipeline of the user's choice, and FFmpeg for the assembly stages.
• Python 3.10+ if you use the Python CLI. Node 18+ if you use the MCP server.
• No accounts. No paid services. No telemetry. No native binaries. No installers.

Data and network

• The Plugin only talks to COMFYUI_URL. It makes no other network calls.
• The Plugin reads workflow JSON files the user names and writes outputs to the user-chosen --output-dir. It does not read or upload model files, LoRAs, voice samples, or any other identity asset.
• The Plugin does not log, persist, or transmit any user data.
• See docs/security-notes.md for the full threat model.

What the Plugin does not do

• It does not train LoRAs. Training is its own project; this Plugin only consumes a LoRA the user already has.
• It does not bundle or distribute any LoRA, model, face embedding, voice sample, or any other identity asset. Every character and voice is user-supplied.
• It does not install packages, run post-install hooks, or download native binaries.
• It does not call any cloud service or third-party API.
• It does not run TTS, edit spreadsheets, burn subtitles, or assemble audio/video. The comfyui-drama Skill documents the full 7-stage pipeline but only ships ComfyUI workflow templates for the two image-side stages (scenarios 5 and 6); the other stages are user pipeline steps that the Plugin deliberately does not assume.

License

Apache-2.0. See LICENSE.