
ComfyUI Builder
io.github.MIt9Published by GitHub account MIt9v2.4.1Updated Oct 24, 2026
ComfyUI MCP server: knowledge base, workflow builder, real-time execution, models, plugins.
Overview
Local MCP server that lets an assistant build, validate and run ComfyUI image-generation workflows, manage models and fetch outputs.
- What you can do
- Discover and inspect nodes, check compatibility and suggest nodes from a bundled knowledge base
- Build workflows programmatically: create, add nodes, connect, validate, plus templates and macros
- Execute workflows with WebSocket progress streaming, batch and chained runs, and progress tracking
- Inspect system resources (GPU/VRAM/RAM) and list or check models across checkpoint, LoRA, VAE and similar types
- List, download and clear outputs and the queue, and interrupt running jobs
- List and reload plugins, and install custom nodes or models through ComfyUI-Manager
- When to use it
- Useful when you want an assistant to turn a plain-language image request into a working ComfyUI graph and run it, without hand-editing JSON. Also fits batch or chained generation and checking hardware limits before building a heavy workflow.
- Requirements
- Runs locally as a Node.js 18+ process over stdio, typically started from a built dist/mcp-server.js with an absolute path. A running ComfyUI instance is needed for execution, queue, model and output tools, configured via COMFYUI_HOST. COMFYUI_PATH is needed for installing custom nodes and models; COMFYUI_KNOWLEDGE_DIR optionally points at the knowledge directory.
Installation
In SourceWeft
- Open ComfyUI Builder in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
mcp-comfy-ui-builder
[Image]Control ComfyUI from Claude or Cursor — 50+ MCP tools for node discovery, dynamic workflow building, real-time WebSocket execution, and model management. No JSON editing required.
[CI] [npm] [Docker] [License: MIT]
Tell your AI assistant what image to generate — it builds and runs the ComfyUI workflow for you.
Works with Claude Desktop, Cursor, and any MCP-compatible client.
Publishable to the MCP Registry via server.json and mcpName.
What is this
- Knowledge base: Seed
knowledge/base-nodes.jsonandnode-compatibility.jsonfrom bundled data (62 seed nodes; 100–600+ after sync). No external services required for suggest/build. - Sync: Custom packs from ComfyUI-Manager; nodes from running ComfyUI via
sync-nodesCLI or on MCP startup. - MCP server (50+ tools): Node discovery, dynamic workflow builder, 9 templates (txt2img, txt2img_flux, img2img, …), WebSocket real-time execution, batch/chain, model management, plugin system.
- Real-time execution: Sub-second progress via WebSocket with automatic polling fallback; ~90% less network traffic for batch runs.
Install
Or from source: git clone https://github.com/MIt9/mcp-comfy-ui-builder.git && cd mcp-comfy-ui-builder && npm install
Quick start
-
Build (postbuild fills knowledge from seed)
-
Use knowledge in code
Commands
Documentation
Single entry point — task-oriented navigation:
- doc/README.md — where to start, task-based navigation
- doc/INDEX.md — complete list of documents and links
- doc/QUICK-REFERENCE.md — commands, examples, troubleshooting
- doc/GETTING-STARTED.md — quick start
- doc/MCP-SETUP.md — connect MCP in Cursor/Claude
- doc/DOCKER-SETUP.md —
docker pull siniidrozd/mcp-comfy-ui-builder| docker-compose for MCP + ComfyUI - Knowledge base: knowledge/README.md, doc/knowledge-base-usage-guide.md
- Workflow Builder: doc/workflow-builder.md — templates, params, save/load, ComfyUI format
- Planning: ROADMAP.md, NEXT-STEPS.md, TODO.md — current v2.3.x, next Phase 10 (QoL)
- Contributing: CONTRIBUTING.md
- MCP Registry: server.json and
mcpNamein package.json; see MCP Registry quickstart to publish afternpm publish.
Requirements
- Node.js 18+
MCP Server (Cursor / Claude)
Server provides 50+ tools across 9 categories:
Core Features
Node Discovery:
- list_node_types, get_node_info, check_compatibility, suggest_nodes
- discover_nodes_live, search_nodes, sync_nodes_to_knowledge
Dynamic Workflow Builder:
- create_workflow, add_node, connect_nodes, validate_workflow
- Build workflows programmatically without JSON manipulation
Templates & Macros:
- list_templates, build_workflow — Pre-built templates (txt2img, img2img, inpainting, upscale, LoRA, ControlNet, batch)
- create_template, apply_template — Parameterized templates
- list_macros, insert_macro — Reusable sub-workflows
Real-time execution 📡
- execute_workflow_sync — Execute with WebSocket progress streaming (sub-second updates)
- execute_workflow_stream — Full event history collection (WebSocket-only)
- get_execution_progress — Real-time progress with node-level granularity
- execute_batch — Concurrent execution with 90% reduced network traffic
- execute_chain — Sequential workflows with data passing
Resources & Model Management:
- get_system_resources — GPU/VRAM/RAM + recommendations (max resolution, model size, batch size); call first before building workflows to avoid OOM
- list_models, check_model_exists, get_workflow_models
- Supports checkpoint, lora, vae, controlnet, upscale, embedding, clip
Outputs & Queue:
- list_outputs, download_output, download_all_outputs
- list_queue, interrupt_execution, clear_queue
Plugins:
- list_plugins, reload_plugins — Data-only plugin system
Install (requires COMFYUI_PATH):
- install_custom_node, install_model — Install nodes and models via ComfyUI-Manager
WebSocket features
- <100ms latency for progress updates (vs 1.5s polling)
- Node-level tracking: See exactly which node is executing with progress percentage
- Automatic fallback: Gracefully falls back to polling if WebSocket unavailable
- Shared connection: Single WebSocket for batch/chain execution (90% reduced traffic)
Set COMFYUI_HOST environment variable for execution/model/output tools:
See doc/MCP-SETUP.md for full tool list and doc/WEBSOCKET-GUIDE.md for real-time features.
Running MCP
From project root, first build the project, then start the server:
Or without npm: node dist/mcp-server.js. Server works via stdio (stdin/stdout). More details → doc/MCP-SETUP.md.
Connect MCP (Cursor / Claude)
What you need: Node.js 18+, one-time build (npm run build), absolute path to dist/mcp-server.js, restart after config change. If the server fails to start with «spawn node ENOENT», use the full path to node in command (run which node and put that path). See doc/MCP-SETUP.md → Troubleshooting.
Minimal config (knowledge and workflow tools only):
Use full path in command (e.g. "/opt/homebrew/bin/node") if Cursor/IDE reports «spawn node ENOENT».
With ComfyUI execution (execute_workflow, get_execution_status, list_outputs, etc.): add "env": { "COMFYUI_HOST": "http://127.0.0.1:8188" } to the server block. See doc/MCP-SETUP.md for full checklist, optional env vars (COMFYUI_PATH, COMFYUI_KNOWLEDGE_DIR), and troubleshooting.
Publish on GitHub
-
Create a new repository on GitHub: name
mcp-comfy-ui-builder, visibility Public (or Private). Do not initialize with README, .gitignore, or license (project already has them). -
Add remote and push:
-
If you fork this repo, replace
MIt9with your GitHub username in the URLs above.
Cursor: Settings → MCP; Claude Desktop: config file ~/Library/Application Support/Claude/claude_desktop_config.json (macOS). Replace the path with your absolute path to dist/mcp-server.js, then restart the app. Full guide: doc/MCP-SETUP.md.
License
MIT
Source: README.md at commit 72928ad
Tools
0Version history
1- v2.4.1LatestSep 16, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts