Comfy API Simplified

This is a small python wrapper over the ComfyUI API. It allows you to edit API-format ComfyUI workflows and queue them programmaticaly to the already running ComfyUI.

I use it to iterate over multiple prompts and key parameters of workflow and get hundreds of images overnight to cherrypick from.

Limitations

Only Basic auth and no auth (for local server) are supported.

Install

pip install comfy_api_simplified          # core library
pip install comfy_api_simplified[mcp]     # + MCP server for AI agents

Use prerequisits

Prepare workflow

You would like to have your node titles unique. Usually both positive and negative prompts have title "CLIP Text Encode (Prompt)", you would like to at least give them different names in case you would like to change it's parameters from python.

Enable "dev options"

In ComfyUI settings, check "Enable Dev mode Options":

Download your workflow in API-format

Have running ComfyUI server

Use

from comfy_api_simplified import ComfyApiWrapper, ComfyWorkflowWrapper

# create api wrapper using your ComfyUI url (add user and password params if needed)
api = ComfyApiWrapper("http://127.0.0.1:8188/")

# create workflow wrapper using your downloaded in api format workflow
wf = ComfyWorkflowWrapper("workflow_api.json")

# change anything you like in your workflow
# the syntax is "Node Title", then "Input param name", then value
wf.set_node_param("Empty Latent Image", "batch_size", 2)
wf.set_node_param("negative", "text", "embedding:EasyNegative")

# queue your workflow for completion
results = api.queue_and_wait_images(wf, "Save Image")
for filename, image_data in results.items():
    with open(f"{filename}", "wb+") as f:
        f.write(image_data)

More examples:

• Queue prompt and get result images example.

• Queue many prompts and do not wait for completion example.

• Send input image and then call i2i workflow example.

MCP Server (for AI agents)

The package ships an MCP server that lets AI agents (Claude, Cursor, etc.) discover nodes and models, build workflows, and run them — all via tool calls.

Start the server

# point at your ComfyUI instance
COMFY_URL=http://192.168.1.x:8188 comfy-mcp-server

# with auth and a custom output directory
COMFY_URL=https://myserver.com COMFY_USER=user COMFY_PASSWORD=pass \
COMFY_OUTPUT_DIR=~/comfy_outputs comfy-mcp-server

The server uses stdio transport (compatible with Claude Desktop, Cursor, and any MCP client).

Available tools

Category	Tool	Description
Discovery	list_node_types	All node class names available in ComfyUI
Discovery	get_node_type_info	Inputs, types, defaults for a node class
Discovery	list_models	Checkpoints, LoRAs, VAEs, etc. grouped by type
Discovery	get_system_stats	GPU VRAM, RAM, Python version
Discovery	list_embeddings	Available text embeddings
Workflow	load_workflow	Load a workflow JSON file into a dict
Workflow	list_nodes	List nodes (id, title, class_type) in a workflow
Workflow	set_node_param	Set a node parameter, returns updated workflow
Workflow	get_node_param	Read a node parameter value
Execution	upload_image	Upload a local image, returns server path
Execution	run_workflow	Run a workflow, save output images, return paths
Execution	get_queue_status	Running/pending queue counts
Execution	interrupt_execution	Stop the current generation

Workflow tools are stateless — the workflow dict is passed in and returned on every call. Chain set_node_param calls to configure a workflow, then pass it to run_workflow.

Add to Claude Desktop

{
  "mcpServers": {
    "comfyui": {
      "command": "comfy-mcp-server",
      "env": { "COMFY_URL": "http://127.0.0.1:8188" }
    }
  }
}

Additional info

There are some other approaches to use Python with ComfyUI out there.

If you are looking to conver your workflows to backend server code, check out ComfyUI-to-Python-Extension

If you are looking to use running ComfyUI as backend, but declare workflow in Python imperatively, check out ComfyScript.

Known issues

If you try to run queue_and_wait_images in async method, it may give you an error since there is already async code inside. As a workaround, you can use

import nest_asyncio
nest_asyncio.apply()

for now.