ComfyUI MCP Server
# ComfyUI_MCP
This repository provides a Model Context Protocol (MCP) server that exposes local ComfyUI workflow files and the remote ComfyUI model catalog to compatible LLM clients.
## Features
- Discovers workflow `.json` files stored under the `workflows/` directory and exposes them as MCP resources.
- Tools for listing workflow metadata and reading workflow contents.
- Tool for recursively querying the ComfyUI `/api/models` endpoints so an LLM can inspect the available checkpoints, LoRAs, and other assets.
- Configuration via CLI arguments or environment variables so the server can be launched from JSON descriptors (e.g., Cursor MCP definitions).
## Installation
Create a virtual environment (recommended) and install the package in editable mode with your preferred installer:
```bash
pip install -e .
```
The project also works with [uv](https://github.com/astral-sh/uv) so you can install or run it without using `pip` directly:
```bash
uv pip install -e .
# or run without installing into the current environment
uvx --from . comfyui-mcp --help
```
## Running the server
The installed `comfyui-mcp` entry point launches the MCP server over stdio. Common configuration options can be supplied either as CLI flags or environment variables:
| Purpose | CLI flag | Environment variable | Default |
|---------|----------|----------------------|---------|
| Workflow directory | `--workflow-dir` | `COMFYUI_WORKFLOW_DIR` | `<repo>/workflows` |
| ComfyUI API base URL | `--api-base-url` | `COMFYUI_API_BASE_URL` | `http://127.0.0.1:8188/` |
| HTTP timeout (seconds) | `--http-timeout` | `COMFYUI_HTTP_TIMEOUT` | `30` |
| Log level | `--log-level` | `COMFYUI_MCP_LOG_LEVEL` | `INFO` |
Example stdio launch configuration for a Cursor MCP JSON definition:
If you prefer `uvx`, the same configuration can be expressed as:
```json
{
"comfyui": {
"command": "uvx",
"args": [
"--from",
".",
"comfyui-mcp",
"--workflow-dir",
"./workflows",
"--api-base-url",
"http://127.0.0.1:8188/"
]
}
}
```
When referencing the project locally with `uvx`, ensure the working directory is set to the
repository root (or adjust the `--from` path accordingly) so the package can be resolved without
requiring it to be published to an external index.
Place your ComfyUI workflow files in the `workflows/` directory (or whatever directory you configure) so they are available to the LLM.
TDQS
Scored across 3 tools
Each tool has a clearly distinct purpose: list_models handles model enumeration, list_workflows lists workflow files, and read_workflow reads specific workflow content. There is no overlap or ambiguity between these functions, making tool selection straightforward for an agent.
All tools follow a consistent verb_noun pattern with snake_case naming: list_models, list_workflows, and read_workflow. This uniformity makes the tool set predictable and easy to understand at a glance.
With only 3 tools, the server feels thin for a ComfyUI integration, which typically involves more operations like executing workflows, managing nodes, or handling images. While the tools are well-defined, the count is borderline low for the apparent scope of interacting with a workflow automation system.
The tool set has significant gaps for a ComfyUI server, missing core operations such as executing workflows, uploading or managing models, and handling generated outputs. This incomplete surface will likely cause agent failures when trying to perform typical ComfyUI tasks beyond basic listing and reading.