Comfy Pilot MCP — Use Cases, Install & Live Demo

Why use it

Key features

Embedded Claude Code terminal (xterm.js) inside ComfyUI
MCP server with direct access to the graph (create nodes, connect edges)
Image analysis from Preview/Save Image nodes
Model downloads from HuggingFace, CivitAI, or direct URLs
Custom node discovery + installation from the registry

Live Demo

What it looks like in practice

comfy-pilot.replay ▶ ready

0/0

Install

Pick your client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "comfy-pilot": {
      "command": "comfy",
      "args": [
        "node",
        "install",
        "comfy-pilot"
      ],
      "_inferred": false
    }
  }
}

Open Claude Desktop → Settings → Developer → Edit Config. Restart after saving.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "comfy-pilot": {
      "command": "comfy",
      "args": [
        "node",
        "install",
        "comfy-pilot"
      ],
      "_inferred": false
    }
  }
}

Cursor uses the same mcpServers schema as Claude Desktop. Project config wins over global.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "comfy-pilot": {
      "command": "comfy",
      "args": [
        "node",
        "install",
        "comfy-pilot"
      ],
      "_inferred": false
    }
  }
}

Click the MCP Servers icon in the Cline sidebar, then "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "comfy-pilot": {
      "command": "comfy",
      "args": [
        "node",
        "install",
        "comfy-pilot"
      ],
      "_inferred": false
    }
  }
}

Same shape as Claude Desktop. Restart Windsurf to pick up changes.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "comfy-pilot",
      "command": "comfy",
      "args": [
        "node",
        "install",
        "comfy-pilot"
      ]
    }
  ]
}

Continue uses an array of server objects rather than a map.

~/.config/zed/settings.json

{
  "context_servers": {
    "comfy-pilot": {
      "command": {
        "path": "comfy",
        "args": [
          "node",
          "install",
          "comfy-pilot"
        ]
      }
    }
  }
}

Add to context_servers. Zed hot-reloads on save.

claude mcp add comfy-pilot -- comfy node install comfy-pilot

One-liner. Verify with claude mcp list. Remove with claude mcp remove.

Use Cases

Real-world ways to use Comfy Pilot

Scaffold an SDXL workflow from a one-line description

👤 ComfyUI users who'd rather describe than drag ⏱ ~20 min intermediate

When to use: You know what you want the workflow to do; you don't want to spend 20 minutes wiring it.

Prerequisites

ComfyUI + comfy-pilot installed — comfy node install comfy-pilot
Claude Code CLI — Pilot installs it automatically if missing

Flow

Open Pilot in ComfyUI

Open the Pilot terminal and say: 'Build me an SDXL text-to-image workflow with ControlNet depth conditioning.'✓ Copied

→ Node graph assembles; models install if missing
Run it

Run the workflow with prompt='cathedral interior, soft light' and my depth map.✓ Copied

→ Image produced; workflow saved

Outcome: A working workflow without manually placing 20 nodes.

Pitfalls

Asking for an exotic model Pilot can't find — Provide a HuggingFace repo URL or CivitAI ID; Pilot downloads whichever is specified

Combine with: comfyui-mcp-server

Refactor an existing messy workflow

👤 Users with workflows that grew organically ⏱ ~25 min intermediate

When to use: Your graph has 60 nodes, half unused, and you want it clean.

Flow

Ask for a cleanup plan

Analyze my current graph; list dead nodes and redundant branches.✓ Copied

→ Claude reads the graph via MCP and produces a plan
Apply

Apply — but keep the Preview Image node at the end.✓ Copied

→ Nodes removed; edges rewired; preview intact

Outcome: Cleaner, easier-to-edit workflow.

Pitfalls

Removing a node that downstream custom code relied on — Have Pilot list touched nodes before destructive ops

Auto-install custom nodes a workflow needs

👤 People opening third-party workflows ⏱ ~10 min beginner

When to use: You imported a workflow and ComfyUI reports missing custom nodes.

Flow

Ask Pilot to resolve

Find and install any custom nodes this workflow needs.✓ Copied

→ Pilot lists missing nodes, then installs from registry
Verify

Reload and confirm the graph no longer has red borders.✓ Copied

→ Graph validates cleanly

Outcome: Unknown workflows usable without spelunking for dependencies.

Pitfalls

Installing unknown custom nodes at scale — Spot-check the package source before letting Pilot install dozens

Combinations

Pair with other MCPs for X10 leverage

comfy-pilot + comfyui-mcp-server

Pilot authors workflows; the MCP server runs them from other clients

Have Pilot build the workflow; then run it from Claude Desktop via comfyui-mcp-server.✓ Copied

comfy-pilot + filesystem

Organize the assets and workflows Pilot creates

After Pilot saves a workflow, move it into ./workflows/<project>/ with a dated name.✓ Copied

comfy-pilot + mcp-image

Compare local ComfyUI results against Gemini generations on the same prompt

Run the same prompt through Pilot's workflow and mcp-image; diff the outputs.✓ Copied

Tools

What this MCP exposes

Tool	Inputs	When to call	Cost
graph.list_nodes / graph.get_node	optional filter	Read the current workflow	0
graph.create_node / graph.connect	type, parent, edges	Author nodes and wire edges	0
image.analyze	node_id (Preview/Save)	Inspect a generated image to decide the next step	0
model.download	source (HF repo, CivitAI id, or URL), target	Missing checkpoints / LoRAs	bandwidth
nodes.install	custom node package(s)	Missing or auto-required custom nodes	bandwidth

Cost & Limits

What this costs to run

API quota: Local compute only
Tokens per call: Small (graph ops); image/model bytes dominate disk and bandwidth
Monetary: Free; your GPU + bandwidth
Tip: Have Pilot confirm before downloading large model bundles; many workflows share checkpoints.

Security

Permissions, secrets, blast radius

Credential storage: Use env vars if you add HF or CivitAI tokens for higher download rates.

Data egress: Downloads from HuggingFace / CivitAI / direct URLs you authorize; otherwise local.

Installing arbitrary custom nodes runs their Python code. Vet the package source before auto-installing at scale.
Direct-URL model downloads can fetch arbitrary files — prefer HF repos and CivitAI IDs with known provenance.

Troubleshooting

Common errors and fixes

Pilot terminal doesn't open

Reinstall via comfy node install comfy-pilot; ensure ComfyUI restart.

Verify: ComfyUI → Manager → Installed, confirm comfy-pilot listed

Custom node install fails behind a corporate proxy

Set HTTPS_PROXY in the ComfyUI process env before launching.

Verify: env | grep -i proxy in the ComfyUI terminal

Model download halts partway

Resume by re-running the download; Pilot resumes where it left off for compatible sources.

Verify: Re-run model.download with the same target path

Alternatives

Comfy Pilot vs others

Alternative	When to use it instead	Tradeoff
comfyui-mcp-server	You want to run workflows from external MCP clients, not author inside ComfyUI	No embedded terminal; different surface
mcp-image	You prefer cloud Gemini over local ComfyUI	Less control; pay per image
everart	You need pre-trained style models	Different provider

More

Resources

📖 Read the official README on GitHub

🐙 Browse open issues

🔍 Browse all 400+ MCP servers and Skills