M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
C

ComfyUI MCP

Community
Drive ComfyUI with natural language.
How to install ↓ Compare GitHub source repository ↗
Category
Other #106 of 230
Stars
★ 838 Very popular
Transport
stdio (local process) · Streamable HTTP
Runtime
Node.js >=22.0.0 · Python
Credentials
No credential needed
License
MIT
Last commit
Tools
44
55FMRS · C

This is a community MCP server and Claude Code plugin maintained by artokun, version 0.5.0, under the MIT license. It offers broad ComfyUI workflow, generation, model, node, diagnostics, and process-control capabilities. stdio is the default transport, with Streamable HTTP also supported.

Strongest · Documentation 14/20 Weakest · Security and permissions 8/20

Reliability
9/20
Security and permissions
8/20
Maintenance
11/20
Documentation
14/20
Setup experience
13/20
Why each score
Reliability 9/20
The README describes stdio startup, HTTP mode, validation, diagnostics, and test scripts, making the main path plausible from static evidence. However, the source files, CI, actual test coverage, and MCP initialization/tool-list handshake cannot be verified. The manifest claims 108 tools and 29 skills while the README claims 181 tools and 35 skills, an unexplained contradiction that warrants a deduction.
Security and permissions 8/20
The default is local stdio; local, remote, and cloud modes are distinguished, credential locations are documented, and some secrets are stated to be excluded from logs. Remote HTTP supports Bearer/X-API-Key authentication, and non-loopback bridge binding requires a token. However, the server can read and write local files, download models, install custom nodes, control processes, call external APIs, and expose remote connections. The material does not adequately verify least privilege, per-operation confirmation, path restrictions, or protection for every dangerous action; explicit HTTP exposure is also available. No evidence of malware, credential theft, or covert exfiltration is present, so the red line is not triggered.
Maintenance 11/20
The MIT license, explicit GitHub repository and maintainer, version number, changelog entry point, contribution-guide entry point, and documented build/test scripts provide a basic maintenance path. The static material contains no verifiable commit history, release cadence, actual CI evidence, security-response channel, or dependency-update evidence, and the maintainer claims cannot be independently verified here, so the score remains moderate.
Documentation 14/20
The README covers installation, client configuration, deployment modes, authentication variables, transports, remote use, tool groupings, examples, troubleshooting, development, and project structure. Deductions remain appropriate because the manifest and README disagree on tool and skill counts, several capabilities are only asserted in the README, cost/privacy/permission boundaries are incomplete, and implementation-level parameters and limits cannot be reviewed.
Setup experience 13/20
The project provides a Node.js requirement, one-command npx startup, a Claude Code configuration example, stdio and HTTP commands, remote ComfyUI options, and cross-platform guidance. It still requires ComfyUI to be installed and running; the Node.js requirement is relatively new; models, custom nodes, credentials, and remote networking add complexity. There is no verifiable CI or end-to-end installation evidence, so the score should not approach full marks.

Static review · not runListed 2026-07-29

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesRuns commands or codeUses the networkConnects to a database

Best for

  • Users who want an agent to operate ComfyUI rather than only relay prompts.
  • Local GPU, LAN, VPS, RunPod, and Comfy Cloud users.
  • Creators who need model-specific generation skills, workflow recipes, and troubleshooting.
  • Users of Claude Code, Claude Desktop, or other MCP-capable agents.

Not for

  • Minimal connector use cases that only forward prompts and return images.
  • Users without a ComfyUI instance or reachable remote ComfyUI.
  • Models that do not support tool calling.
  • Users seeking official Comfy maintenance or an official cloud-service guarantee.

Required permissions

  • Requires access to the ComfyUI REST API and WebSocket.
  • Local mode may read and write model, input, output, and workflow-library directories.
  • Process-control features may start, stop, or restart local ComfyUI.
  • Model and node discovery or downloads may access Hugging Face, CivitAI, GitHub, and the ComfyUI Registry; some features require tokens.
  • Comfy Cloud mode requires COMFYUI_API_KEY.

Risks and side effects

  • Workflow execution can consume local GPU VRAM, compute, and disk space.
  • Downloaded models and custom nodes can be large and may include third-party code.
  • stop_comfyui, restart_comfyui, and queue operations can interrupt jobs or unsaved workflow changes.
  • Remote HTTP, tunnel, or bridge deployments increase network exposure and should use tokens and access controls.
  • Hosted API nodes may consume paid API credits.

Setup

Before you start

Runtime：Node.js >=22.0.0 · Python

COMFYUI_MCP_BRIDGE_TOKEN optionalsecret Shared secret gating the bridge WebSocket; mandatory for non-loopback binds.
COMFYUI_API_KEY optionalsecret Comfy Cloud API key; when set, cloud mode targets cloud.comfy.org.
COMFYUI_AUTH_TOKEN optionalsecret Auth token for a self-hosted ComfyUI behind a reverse proxy.
CIVITAI_API_TOKEN optionalsecret CivitAI token for model downloads, from your civitai.com account settings.
HUGGINGFACE_TOKEN optionalsecret HuggingFace token for higher API rate limits, from huggingface.co.
GITHUB_TOKEN optionalsecret GitHub token to avoid rate limits when generating node skills.
REGISTRY_ACCESS_TOKEN optionalsecret Comfy Registry API key used only by publish_custom_node; env-only.
Other optional settings (23)
COMFYUI_URL optional Full ComfyUI URL for a local or remote instance; overrides auto-detection.
COMFYUI_HOST optional ComfyUI server address, defaults to 127.0.0.1.
COMFYUI_PORT optional ComfyUI server port; auto-detects 8188 then 8000 by default.
COMFYUI_PATH optional Path to the ComfyUI data directory, set if auto-detection fails.
COMFY_CLI_PATH optional Path to the official comfy executable (requires comfy-cli >=1.11.1).
COMFYUI_PYTHON optional Python interpreter for legacy dependency-install fallbacks, defaults to python.
COMFYUI_MCP_BRIDGE_HOST optional Panel-orchestrator bridge bind host; non-loopback requires a token.
COMFYUI_MCP_DATA_DIR optional Per-instance data dir (incl. generations.db), defaults to ~/.comfyui-mcp.
COMFYUI_CLOUD_URL optional Overrides the Comfy Cloud endpoint, default https://cloud.comfy.org.
COMFYUI_AUTH_HEADER optional Header name for the auth token, default Authorization (e.g. X-API-Key).
COMFYUI_AUTH_SCHEME optional Scheme prefix on the token value (e.g. Bearer, Token).
COMFYUI_DOWNLOAD_CACHE_DIR optional Content-addressed model download cache dir, default ~/.comfyui-mcp/cache.
COMFYUI_LRU_CACHE_SIZE_GB optional LRU cap on the download cache in GB; 0 disables eviction.
COMFYUI_STARTUP_CHECK_INTERVAL_S optional Readiness-probe interval in seconds when starting local ComfyUI, default 1.
COMFYUI_ALWAYS_RESTART optional Auto-restart a crashed local ComfyUI, default false.
COMFYUI_MCP_STALL_S optional Render-wedge watchdog seconds, default 180 (clamped 15-3600).
COMFYUI_MCP_INTERRUPT_S optional Seconds cancel_job waits for an interrupt before escalating, default 30.
COMFYUI_MCP_FORCE_REMOTE optional Set 1/true to force remote mode for loopback URLs (e.g. port-forwarded RunPod).
COMFYUI_MCP_TOOL_MODE optional Set to compact to register 3 meta-tools instead of the full schema surface.
MCP_TRANSPORT optional Set to http to serve streamable-HTTP at /mcp instead of stdio.
MCP_HOST optional HTTP bind host, default 127.0.0.1.
MCP_PORT optional HTTP port, default 9100.
LOG_LEVEL optional Logging verbosity: debug/info/warn/error, default info.

Install ComfyUI, then add the server configuration from server.json to an MCP client. For Claude Code, place command npx and args ["-y","comfyui-mcp"] under mcpServers.comfyui in ~/.claude/settings.json. Start ComfyUI and request generation or workflow operations in natural language. Streamable HTTP is available with npx -y comfyui-mcp@latest --http; remote instances can be selected with --comfyui-url.

claude_desktop_config.json 
{"mcpServers":{"comfyui":{"command":"npx","args":["-y","comfyui-mcp"],"env":{"CIVITAI_API_TOKEN":""}}}}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "comfyui": {
      "command": "npx",
      "args": [
        "-y",
        "comfyui-mcp"
      ],
      "env": {
        "CIVITAI_API_TOKEN": ""
      }
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add comfyui -e CIVITAI_API_TOKEN= -- npx -y comfyui-mcp

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

With ComfyUI running, ask your client 'What are my ComfyUI system stats?' — if the get_system_stats tool returns GPU/VRAM info, the server is connected and working.

Troubleshooting

  1. Confirm that ComfyUI is running; the server probes ports 8188 and 8000 by default.
  2. Set COMFYUI_PATH if installation-path auto-detection fails.
  3. Use --comfyui-url or COMFYUI_URL for remote instances; use --force-remote for port-forwarded remote installs.
  4. Check HUGGINGFACE_TOKEN or CIVITAI_API_TOKEN when model downloads fail.
  5. Use get_history, get_logs, or /comfy:debug to inspect node, model, and Python errors.
  6. Use clear_vram and check model VRAM requirements when memory is insufficient.
  7. Ensure the selected model supports tool calling; the documentation also strongly recommends thinking and vision capabilities.

Things to try

Once connected, you can ask your AI assistant things like:

  • Generate an image of a sunset over mountains
  • Free my VRAM
  • Visualize this workflow as a Mermaid diagram
  • What checkpoints do I have installed?

Tools 44

generate_image writes
Generate an image from a text prompt and build a workflow automatically.
generate_with_controlnet writes
Generate an image using a ControlNet image condition and prompt.
generate_with_ip_adapter writes
Generate an image guided by a reference image's style or subject.
generate_audio writes
Generate music, instruments, or sound effects from a text prompt.
view_image read-only
Return a generated asset as an inline image for agent inspection.
regenerate writes
Re-run the workflow that produced an asset, with optional parameter overrides.
list_assets read-only
Browse recently generated assets by time.
get_asset_metadata read-only
Return full asset provenance, workflow, and metadata.
Show 36 more tools
get_defaults read-only
Show merged generation defaults with source attribution.
set_defaults writes
Update runtime generation defaults and optionally persist them.
enqueue_workflow writes
Submit an API-format workflow and return a prompt ID.
get_job_status read-only
Check the execution status of a job.
get_queue read-only
View running and pending jobs in the queue.
cancel_job destructive
Interrupt a running job and optionally clear pending jobs.
get_system_stats read-only
Return GPU, VRAM, Python, and operating-system information.
visualize_workflow read-only
Convert a workflow into a Mermaid flowchart grouped by category.
mermaid_to_workflow read-only
Convert a Mermaid diagram into executable workflow JSON.
create_workflow writes
Create txt2img, img2img, upscale, inpaint, and related workflows from templates.
modify_workflow writes
Set inputs, add or remove nodes, and connect nodes.
get_node_info read-only
Query node types available in ComfyUI.
validate_workflow read-only
Check nodes, connections, output indices, and model files without executing.
list_workflows read-only
List saved workflows in ComfyUI's user library.
get_workflow read-only
Load a saved workflow by filename.
strip_workflow read-only
Expand virtual links, reroutes, subgraphs, and bypassed nodes into a flat graph.
slice_workflow read-only
Extract one standalone enabled pipeline from a toggle-template workflow.
save_workflow writes
Save a workflow to ComfyUI's user library.
upload_image writes
Copy a local image into ComfyUI's input directory.
workflow_from_image read-only
Extract workflow metadata from a ComfyUI-generated PNG.
list_output_images read-only
Browse recent images and videos in the output directory.
search_models read-only
Search Hugging Face for compatible models.
download_model writes
Download a model URL to the appropriate ComfyUI subdirectory.
list_local_models read-only
List installed local models by type.
clear_vram writes
Free GPU VRAM through ComfyUI's endpoint.
get_embeddings read-only
List installed textual-inversion embeddings.
search_custom_nodes read-only
Search the ComfyUI Registry for custom node packs.
get_node_pack_details read-only
Return a node pack's description, author, nodes, and installation information.
generate_node_skill writes
Generate node-pack skill documentation from a Registry ID or GitHub URL.
get_logs read-only
Return ComfyUI server logs with optional keyword filtering.
get_history read-only
Return execution history, errors, traces, and timing details.
stop_comfyui writes
Stop the running ComfyUI process.
start_comfyui writes
Start ComfyUI using saved process information.
restart_comfyui writes
Restart ComfyUI while preserving launch arguments.
suggest_settings read-only
Suggest sampler, scheduler, steps, and CFG settings from local generation history.
generation_stats read-only
Show local generation tracking statistics.

Use cases

Generate images, video, and audio from prompts.
Create, modify, validate, and execute ComfyUI workflows through an agent.
Search, download, and manage models and custom nodes.
Diagnose workflow failures using logs, history, and VRAM information.
Load, slice, and flatten complex workflows on the live canvas.

Supported clients

Claude Desktop
Claude Code
MCP Inspector
HermesPartial support
OpenClawPartial support
GitHub Copilot CLIPartial support
OllamaPartial support

Listed from the project's documentation, not tested by this site.

Overview

comfyui-mcp is a local-first, agent-native control plane for ComfyUI, distributed as an MCP server and Claude Code plugin. It generates images, video, and audio; creates, executes, validates, visualizes, and modifies workflows; manages models, embeddings, VRAM, and custom nodes; and edits the live ComfyUI graph. It supports local installs, LAN, VPS, remote ComfyUI, and Comfy Cloud.

Similar servers

BioMCP 75 · B

One binary. One grammar. Evidence from the biomedical sources you already trust.

★ 659 · Tools 11 Compare with this →
Compare with other MCP servers

Source revision 6996d06ff511 Data synced 2026-10-19 Read the FMRS scoring method