comfyui-mcp — drive ComfyUI with ANY LLM

[!IMPORTANT]

This project is no longer maintained.

ComfyUI now ships official agent and MCP tooling — [Comfy Agent](https://comfy.org/agent) and

[Comfy MCP](https://comfy.org/mcp) — built and supported by the Comfy-Org team with deeper

integration than a community project can match. If you’re looking for an MCP server or AI agent

for ComfyUI, use the official tooling.

This repo will remain public as a reference, but no new features, bug fixes, or dependency

updates will be made. Issues and pull requests close when the repo is archived on 2026-10-09.

The community Discord goes read-only the same day, with its threads left up as a searchable

archive.

Thanks to everyone who used, starred, forked, and contributed to this project.

It was a good run. :rocket:

The full story is in the goodbye post.

Looking for a ComfyUI or generative-AI integration expert? I take on custom solutions — reach me by

email at art.longbottom.jr@gmail.com or on LinkedIn.

<p align="center">

<a href="https://comfyui-mcp.artokun.io/docs">

<img src="docs/images/demo-poster.jpg" width="760"

alt="The Agent panel driving ComfyUI end to end. Click to watch the demo">

</a>

</p>

<p align="center">

<em>The Agent panel driving ComfyUI end to end. It reads what is installed locally,

wires the graph, frees VRAM, and runs the render.<br>

<a href="https://comfyui-mcp.artokun.io/docs">Watch the 76s demo &rarr;</a></em>

</p>

The local-first, agent-native control plane for [ComfyUI](https://github.com/comfyanonymous/ComfyUI). An MCP server plus a live sidebar agent that generates images, video and audio, authors and runs workflows, manages models and custom nodes, and edits your live ComfyUI graph in natural language. Bring whatever model you have: Claude or ChatGPT on your subscription, Gemini on your Google login, a free local model via Ollama (fully offline), or any hosted model over one API key (DeepSeek, GLM, MiMo, Kimi, GPT, Claude via OpenRouter). The tools and the panel are the same on every tier, and the built-in LLM Arena scores each model on real ComfyUI tasks so you know what yours can do. One config targets local installs, LAN, VPS, or Comfy Cloud.







  &nbsp;One-click GPU pod. A ready-to-run ComfyUI with this project + Agent Panel + ComfyUI-Manager v2 preinstalled, on your own GPU. No setup.

Works on macOS, Linux, and Windows. Auto-detects your ComfyUI installation and port.

Stuck or have a question? [Join the Discord](https://discord.gg/cW9arBhzCu) for help, model tips, and release announcements.

38 MCP tools | 42 AI skills (Flux · WAN · LTX 2.3 video · MiniMax H3 · Qwen · Z-Image · Ideogram 4 · ERNIE · ANIMA · model registry · Civitai · node authoring · launch/perf flags) | 56 installer packs | 11 slash commands | 4 autonomous agents | 3 hooks

The plugin ships skills that grow with every release: model-specific generation guides with curated download URLs, workflow recipes, troubleshooting, and custom-node authoring. With them, Claude knows the right sampler, CFG, resolution, and model files for each architecture without trial and error.

### Now available: the ComfyUI Agent Panel on ComfyUI-Manager & the Comfy Registry

An autonomous AI agent in your ComfyUI sidebar, running on Claude, ChatGPT, Gemini, or ANY local/hosted LLM (Ollama and every OpenAI-compatible endpoint). Subscriptions work with no API key; local models work with no account at all. Pick a provider and it drives your live graph: edits, spatial layout, one-shot workflow/pack loads, rewind/rollback, a pending-message tray, activity cards, multi-tab. It asks before spending paid API credits.

Search comfyui-agent-panel in ComfyUI-Manager to install. Read more →

Full documentation: [comfyui-mcp.artokun.io/docs](https://comfyui-mcp.artokun.io/docs)

Quick Start

1. Install ComfyUI (if you haven't already): ComfyUI Desktop or from source

2. Add the MCP server to your Claude Code config (~/.claude/settings.json):

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

3. Start using it. With ComfyUI running, ask Claude to generate an image:

> Generate an image of a sunset over mountains

Claude will find (or download) a checkpoint, build a workflow, execute it, and return the image.

Note: This runs as a standalone MCP server, so there is no need to clone this repo. npx downloads and runs it.

Scope: local, remote, or Comfy Cloud

comfyui-mcp is local-first. A self-hosted ComfyUI on Mac/Linux/Windows is the primary target, and the same agent reaches remote installs (RunPod, VPS, LAN, reverse-proxied) from one config. Local-first, not local-only.

More than a bridge. Most ComfyUI MCP servers are thin connectors. They forward a prompt and hand back an image. comfyui-mcp is a full control plane. It authors and edits the graph node by node, runs and iterates on workflows, manages models and custom nodes, and ships model-specific expertise (samplers, CFG, resolutions, curated model URLs) so the agent gets it right without trial and error. If you want a minimal local relay, a lightweight server is fine. If you want an agent that operates ComfyUI, that's this project.

For Comfy Cloud users, Comfy-Org ships its own agent tooling: Comfy Cloud MCP (public beta, hosted on Comfy Cloud GPUs), the Comfy In-App Agent (private alpha, inside Comfy Cloud), and a first-party Comfy Local MCP (private test, not publicly available yet), all maintained by the Comfy team. If you don't have a GPU or you want zero setup, that's the better path; go use it.

Where this project differs is that it runs on your install and your choice of model, including a free local one via Ollama with no account and no network at all. comfyui-mcp also includes a community cloud mode (set COMFYUI_API_KEY, see Deployment modes) so a single MCP can target all three deployment shapes from one config.

[Local vs. Comfy Cloud agent](https://comfyui-mcp.artokun.io/docs/local-vs-comfy-cloud) is a side-by-side comparison, including when Comfy Cloud is the right answer. (Statuses above are as of July 2026. Comfy-Org ships fast, so check their docs for the current state.)

Remote / hosted connector (one command)

To use comfyui-mcp from Claude Desktop's Custom Connectors or any remote

client, like Comfy's own cloud.comfy.org/mcp connector, run it as an

authenticated, publicly reachable Streamable-HTTP server with one flag:

npx -y comfyui-mcp@latest --tunnel

This forces the HTTP transport, generates an auth token, opens a

cloudflared quick tunnel, and prints

a ready-to-paste https://…/mcp URL + token + Claude Desktop connector snippet.

Auth accepts Authorization: Bearer <token> or X-API-Key: <token> (matching

Comfy Cloud's convention). See the

Remote / hosted connector guide

for the full walkthrough and headless usage.

Auth is opt-in. With no COMFYUI_MCP_HTTP_TOKEN set and no --tunnel, the

default stdio (and plain --http on loopback) behavior is unchanged. It stays open and

local. OAuth (Comfy's browser sign-in flow) is a planned follow-up.

Claude Code Plugin

This package also ships as a Claude Code plugin that adds slash commands, skills, agents, and hooks on top of the MCP tools.

Install as a plugin

# In Claude Code
/plugin marketplace add artokun/comfyui-mcp
/plugin install comfy

Slash commands

| Command | Description |

|---------|-------------|

| /comfy:gen <prompt> | Generate an image from a text description — auto-selects checkpoint, builds workflow, returns image |

| /comfy:viz <workflow> | Visualize a workflow as a Mermaid diagram with nodes grouped by category |

| /comfy:node-skill <pack> | Generate a Claude skill for a custom node pack from Registry ID or GitHub URL |

| /comfy:debug [prompt_id] | Diagnose why a workflow failed — reads history, logs, traces root cause, suggests fixes |

| /comfy:batch <prompt, params> | Parameter sweep generation across cfg, sampler, steps, seed, etc. |

| /comfy:convert <file> | Convert between UI format and API format workflows |

| /comfy:install <pack> | Install a custom node pack — git clone, pip install, optional restart |

| /comfy:gallery [filter] | Browse generated outputs with metadata — filter by date, count, or filename |

| /comfy:compare <a vs b> | Diff two workflows side by side — shows added/removed nodes and changed parameters |

| /comfy:recipe <name> <prompt> | Multi-step recipes: portrait, hires-fix, style-transfer, product-shot |

Built-in skills

42 skills total: model-family guides (Flux, WAN, LTX 2.3, MiniMax H3, Qwen, Z-Image, Ideogram 4, ERNIE, ANIMA + anime / WAN / Z-Image LoRA training), the model-registry (curated download URLs), the civitai pairing skill, node authoring, the launch/performance-flags matrix, and the core four below. Full list on the plugin docs page.

Installer packs. `packs/` bundles 13 one-command ComfyUI setups: ANIMA, Ideogram 4, LTX-2.3, ERNIE, WAN (animate / longer-videos / transparent), Qwen (image / image-edit), Z-Image (turbo / base / xy-plot) and artokun-flow (WAN Animate: replace / animate). Each is a manifest of custom nodes + model URLs + workflow that drives both apply_manifest and generated install-windows.bat / install-runpod.sh, with CI that validates every model link + payload size. See `packs/README.md`.

| Skill | Description |

|-------|-------------|

| comfyui-core | Workflow format, node types, data flow patterns, pipeline architecture, MCP tool usage guide |

| prompt-engineering | CLIP weight syntax (word:1.3), BREAK tokens, embeddings, model-specific prompting for SD1.5/SDXL/Flux/SD3 |

| troubleshooting | Common error catalog — OOM, dtype mismatches, missing nodes, NaN tensors, black images, CUDA errors, with VRAM estimates per model |

| model-compatibility | Compatibility matrix — loaders, resolutions, CFG, samplers, ControlNets, LoRAs, and VAEs per model family (SD1.5/SDXL/Turbo/Lightning/Flux/SD3/LTXV) |

Agents

| Agent | Model | Description |

|-------|-------|-------------|

| comfy-explorer | Sonnet | Researches custom node packs — reads docs, queries /object_info, generates skill files |

| comfy-debugger | Sonnet | Autonomously diagnoses workflow failures — gathers logs + history, identifies failing node, checks models + custom nodes, proposes and optionally applies fixes |

| comfy-optimizer | Sonnet | Analyzes workflows for performance — detects redundant nodes, VRAM waste, wrong CFG/steps for model family, precision issues, suggests optimizations |

| comfy-researcher | Sonnet | Discovers and ranks ComfyUI custom node packs for a stated image-generation problem |

Hooks

| Event | Trigger | Action |

|-------|---------|--------|

| PreToolUse | enqueue_workflow | VRAM watchdog — checks GPU memory via /system_stats and warns if < 1GB free before execution |

| PreToolUse | restart_comfyui (actions stop/restart) | Save warning — prompts user to save unsaved workflow changes before stopping ComfyUI |

| PostToolUse | Any comfyui tool | Job completion notify — checks for completed jobs and injects completion summaries into the conversation |

Background Scripts

| Script | Description |

|--------|-------------|

| monitor-progress.mjs | Progress monitor — connects to ComfyUI's WebSocket for real-time step progress (e.g., step 5/14 (36%)). Run as a background Bash task after enqueuing workflows. Reports completion with output filenames, errors with node details. Replaces polling queue (action:"status") in a loop. |

Panel agent (Claude · ChatGPT · Gemini · any local/hosted LLM)

Beyond the headless MCP server, this package ships the panel orchestrator that

powers the ComfyUI Agent Panel,

an autonomous agent embedded in ComfyUI's sidebar that drives the live canvas. It

runs in the background on your own subscription (Claude or ChatGPT), started

on demand by the panel's Connect button:

npx -y comfyui-mcp@latest connect

Drive a REMOTE ComfyUI from your own machine (connect)

When ComfyUI runs somewhere with no Node or agent (a RunPod pod, a cloud box) you

can still run the agent on your machine and drive that remote ComfyUI. There is no

agent login on the box and nothing to install or configure remotely:

npx -y comfyui-mcp@latest connect https://abcd1234-3000.proxy.runpod.net

This is sugar for --panel-orchestrator with COMFYUI_URL set from the URL. The

orchestrator runs locally on your Claude/ChatGPT login and reaches the remote

ComfyUI over its public proxy URL. For a remote HTTPS pod, connect

opens a token-gated wss:// tunnel (via Cloudflare) to

the local agent bridge and hands the pod's panel that URL, so the pod's HTTPS page

reaches your machine with no browser prompt, in any browser. A secure page

can't open a plain ws:// socket to your box because of mixed-content and Private Network

Access rules. A local ComfyUI uses the plain ws://127.0.0.1:9199 loopback bridge.

Add --insecure-bridge to force that loopback for a remote pod, then arrange

your own path to it, e.g. an SSH port-forward. Either way the panel JS runs in

your local browser, and the agent and your login run only on your

machine, so nothing is installed remotely.

To finish, with connect still running on your machine, open the remote ComfyUI in

your browser and click Connect in the Agent panel. The panel is a pure-frontend

extension. It links to the bridge your connect process is already serving, rather

than asking the ComfyUI host to spawn an orchestrator it cannot run.

Multi-provider, full parity. The orchestrator depends on a provider-neutral

AgentBackend port (dependency injection), with two adapters:

• ClaudeBackend wraps the Claude Agent SDK

(@anthropic-ai/claude-agent-sdk), a persistent streaming session over the

claude.ai subscription (OAuth, no key).

• CodexBackend wraps OpenAI Codex over the codex app-server JSON-RPC

protocol (@openai/codex), on the ChatGPT subscription (codex login, no key).

Further adapters follow the same port: Gemini CLI (ACP), Antigravity

(agy, the Google AI Pro/Ultra subscription path; install from

antigravity.google and run agy once to sign

in), Grok, Kimi, GLM, Ollama/LM Studio/llama.cpp (local), OpenRouter, and

any OpenAI-compatible endpoint. See

docs/backends for the full

matrix.

Both are optional dependencies, and the panel picks a provider, not a port.

Each backend runs its own orchestrator on its own loopback bridge port. A

capability matrix tells the panel which features each backend supports (e.g. conversation-rollback is

Claude-only today, since the Codex app-server resumes whole threads only).

The live-canvas tools and model knowledge are identical across providers. The

panel_* tool definitions live in one shared list, registered onto both the

in-process Claude SDK MCP server and a @modelcontextprotocol/sdk server over a

loopback streamable-HTTP MCP that the orchestrator hosts for Codex (which can

only host config-declared MCP servers). The headless comfyui MCP is likewise

injected into both, in-process for Claude and declared via `codex app-server -c

mcp_servers` for ChatGPT, so generation, models, and workflow tools are the same

everywhere.

New tools that give every backend the same expertise and a cost guardrail:

| Tool | Description |

|------|-------------|

| list_packs (action: "skill_list" / "skill_read") | Discover and read bundled model-family + workflow skills — the knowledge Claude loads natively, exposed to any MCP client (e.g. the Codex backend) |

| list_packs (action: "list" / "read_workflow") | List one-command installer packs (custom nodes + weights + ready workflow; all local-GPU / free) and read a pack's graph |

| list_packs (action: "list_templates") | List the connected ComfyUI's custom-node-contributed workflow templates |

| list_packs (action: "check_runtime") | Classify a workflow as local (your GPU, free) or api / mixed / unknown (hosted API nodes = paid credits) so the agent asks before spending paid API credits |

| list_packs (action: "extract_deps" / "install_deps") | Work out which custom node packs a workflow needs, and install the missing ones through ComfyUI-Manager |

| panel_load_workflow | (panel tool) Load a full workflow onto the live canvas in one shot — by bundled pack name (read server-side, never shuttled through chat) or by graph JSON |

| panel_strip_workflow / panel_slice_workflow | (panel tools) De-virtualize a tangled graph (Get/Set buses, Reroutes, subgraphs, bypass → real connections) or carve one rgthree-toggled pipeline out of a monolith — by pack, server-side path, or inline graph; for understanding/rebuilding expert workflows without hand-tracing |

MCP Tools

38 tools across workflow execution, generation, iteration, composition, models, and more:

Image Generation (high-level)

| Tool | Description |

|------|-------------|

| generate_image action: "image" | Generate from a text prompt — builds a txt2img workflow, fills unspecified params from your defaults, auto-selects a checkpoint |

| generate_image action: "controlnet" | Generate conditioned by a ControlNet image (pose/depth/canny/normal) + prompt |

| generate_image action: "ip_adapter" | Generate guided by a reference image's style/subject via IP-Adapter (needs ComfyUIIPAdapterplus) |

| generate_image action: "video" / "3d" | Generate a short video clip (LTX-2.3, local GPU) or a 3D model (hosted partner API nodes) from the same one-line entry point |

| generate_image action: "upscale" | Post-process an uploaded or staged image with ESRGAN super-resolution |

| generateimage (action:"removebackground") | Post-process an uploaded or staged image into a transparent BiRefNet cutout |

Audio Generation (high-level)

| Tool | Description |

|------|-------------|

| generate_image action: "audio" | Generate audio from a text prompt — supports ACE Step 1.5 (music with lyrics/structure) and Stable Audio 3 (music, instruments, SFX); auto-selects local models |

Assets & Iteration

| Tool | Description |

|------|-------------|

| get_image (action:"view") | Return a generated asset's bytes as an inline image so the agent can see the result |

| get_image (action:"analyze_color") | Palette / contrast / color statistics for a generated image (dominant colors, average + luminance stats, contrast checks) so the agent can reason about color without a vision round-trip |

| generateimage (action:"regenerate") | Re-run the workflow that produced an `assetid`, with optional parameter overrides |

| get_image (action:"list_assets") | Browse recently generated assets (newest-first) by asset_id |

| get_image (action:"asset_metadata") | Full provenance for an asset, including the originating workflow |

Defaults

| Tool | Description |

|------|-------------|

| get_defaults action:"get" | Show merged generation defaults with per-source attribution |

| get_defaults action:"set" | Update runtime defaults; persist: true writes the config file |

| get_defaults action:"get_ui" / action:"set_ui" | Read/write ComfyUI's OWN frontend UI settings (the Comfy.* ids) — a separate store from the generation defaults above |

Workflow Execution

| Tool | Description |

|------|-------------|

| enqueue_workflow action: "enqueue" | Submit a workflow (API format JSON) — returns prompt_id immediately, non-blocking |

| enqueue_workflow action: "rerun" / "run_url" / "template_schema" | Re-run a past generation, or read/run a shared workflow from a URL, or inspect a bundled template's overridable slots |

| enqueueworkflow (action:"runtemplate") | One-shot: resolve a bundled pack's expert graph, apply <nodeId>.<widget> overrides, and enqueue it |

| queue | One action-parameterized tool for the execution queue: list (running + pending), status (one job by prompt ID), get_workflow (a pending job's full payload), move/edit (requeue a pending job front/back, patched or replaced, with a new prompt ID), cancel (interrupt the running job — escalates interrupt → verify → /free, reports WEDGED if it won't die; clear_pending: true also drops all pending), cancel_queued, clear |

| get_system_stats | Get system info — GPU, VRAM, Python version, OS |

Workflow Visualization

| Tool | Description |

|------|-------------|

| visualize_workflow | One action-parameterized tool for rendering and converting a workflow you pass in: render (Mermaid flowchart, nodes grouped by category), render_hierarchical (the same graph sectioned — overview, one section in detail, a listing, or an AI-oriented summary), mermaid (a Mermaid diagram back to executable workflow JSON), to_dsl/from_dsl (the compact, losslessly round-tripping authoring DSL) |

Workflow Composition

| Tool | Description |

|------|-------------|

| create_workflow | One action-parameterized tool for authoring: create (from templates: txt2img, img2img, upscale, inpaint, controlnet, ip_adapter, ace_step_15, stable_audio_3), modify (operations: set_input, add_node, remove_node, connect, insert_between), validate (dry-run — missing nodes, broken connections, invalid output indices, missing model files), node_info (query available node types from ComfyUI's /object_info endpoint) |

Workflow Library

| Tool | Description |

|------|-------------|

| get_workflow | One action-parameterized tool for READING a saved workflow file: list (the user library, subfolders included), get (one workflow's JSON by filename), analyze (a structured summary instead of raw JSON), query (filter/traverse/aggregate a big graph without dumping it), strip (de-virtualize any workflow from an absolute path, library filename, or inline graph — resolve GetNode/SetNode buses, Reroutes, subgraph defs, and bypassed nodes into real connections and return the flat graph; reads ANY path server-side, so it loads ad-hoc/expert workflows the cached library can't), slice (un-chunk a toggle-template monolith — one rgthree Fast-Groups-Bypass-toggled pipeline out into a standalone activated graph; pair with strip to then flatten the buses), from_image (the workflow ComfyUI embedded in a PNG), prompt_director (Prompt Director's sanitized runtime state) |

| save_workflow | One action-parameterized tool for WRITING to the library: save (store a workflow — overwrites a same-filename file), lock (record a provenance lock: SHA-256 per model, git commit per node pack), verify_lock (report drift against that lock) |

Image Management

| Tool | Description |

|------|-------------|

| upload_image | Copy a local image into ComfyUI's input/ directory for img2img, inpaint, or ControlNet |

| get_image (action:"list_outputs") | Browse recently generated images and videos from the output directory, sorted newest-first — recurses into subfolders (e.g. SaveVideo's output/video/…) and returns each result's subfolder |

Model Management

| Tool | Description |

|------|-------------|

| download_model | Find and fetch model weights, and track the transfers. action: download (from a URL, into the correct ComfyUI subdirectory), status, cancel, search (HuggingFace), search_civitai, search_creators, download_civitai, resolve_missing |

| list_local_models | What is installed, and where ComfyUI looks. action: list (installed models by type: checkpoints, loras, vae, upscalemodels, controlnet, embeddings, clip, unet, diffusionmodels, text_encoders), remove (deletes a model file), embeddings, list_paths, add_path, remove_path (the last three view/edit the extra search-path YAML) |

Memory Management

| Tool | Description |

|------|-------------|

| clear_vram | Free GPU VRAM by unloading cached models — calls ComfyUI's /free endpoint, reports before/after stats |

Registry & Discovery

| Tool | Description |

|------|-------------|

| search_custom_nodes | Search the ComfyUI Registry for node packs by keyword (action: "search"), or get one pack's full details (action: "details") |

| list_packs (action: "generate_skill") | Generate a Claude skill .md file from a Registry ID or GitHub URL |

| comfy_cli | Search actual loaded node classes through official comfy nodes search (action:"search_nodes") |

Official comfy-cli

Install comfy-cli 1.11.1 or newer to enable the official JSON-backed tools. The MCP resolves comfy from COMFY_CLI_PATH, PATH, or the selected workspace's .venv/venv. Local custom-node operations prefer comfy node when a supported CLI is available and otherwise fall back to Manager HTTP; remote targets keep using Manager HTTP.

| Tool | Description |

|------|-------------|

| comfy_cli | One action-parameterized tool for the whole official CLI: status, server_start/server_stop/server_restart, jobs_list/jobs_status/jobs_wait/jobs_watch/jobs_cancel, search_nodes, workflow_validate/workflow_run, transfer_upload/transfer_download, models_* (list/search/show/download/remove), skills_* (list/show/validate/install/status/uninstall) |

Diagnostics

| Tool | Description |

|------|-------------|

| get_system_stats (action:"logs") | Get ComfyUI server logs with optional keyword filter (e.g., error, warning, a node name) |

| kitchen | See what comfy-kitchen can do on this GPU (status), find where a graph leaves it on the table (assess), and apply the faster path (apply) — a failed probe is unknown, never a no |

| get_history action: "list" | Get execution history with full error details, Python tracebacks, timing, and cached node info |

| get_history action: "diagnose" | Explain a FAILED run in one call — the failed node and traceback PLUS the missing models (file + widget) and missing node types |

Process Control

| Tool | Description |

|------|-------------|

| restart_comfyui action:"restart" | Stop and restart ComfyUI, preserving all launch arguments |

| restart_comfyui action:"stop" | Stop the running ComfyUI process (saves PID and launch args for restart) |

| restart_comfyui action:"start" | Start ComfyUI using info saved from a previous stop |

Generation Tracker

| Tool | Description |

|------|-------------|

| get_history action: "suggest" | Suggest proven sampler/scheduler/steps/CFG settings from local generation history — query by model family, LoRA hash, or text search |

| get_history action: "stats" | Show local generation tracking statistics — total runs, unique combos, breakdown by model family |

Every enqueue_workflow call logs settings to a local SQLite database (generations.db). Repeated settings combos get a reuse_count bump instead of a duplicate row, so the count works as a popularity signal. The tracker identifies models and LoRAs by content hash (AutoV2 / SHA256), not filename, so renamed files still group together.

# View local stats from the CLI
npm run generations:stats

Examples

Generate an image

> /comfy:gen a cyberpunk city at night with neon lights

Claude will:

1. Check installed checkpoints (download one if needed)
2. Build a txt2img workflow with your prompt
3. Execute it on ComfyUI
4. Return the generated image

Visualize a workflow

> /comfy:viz ~/workflows/my-workflow.json

Produces a Mermaid diagram with nodes grouped by category:

flowchart LR
  subgraph Loaders
    1["CheckpointLoaderSimple"]
  end
  subgraph Conditioning
    2(["Positive Prompt"])
    3(["Negative Prompt"])
  end
  subgraph Sampling
    5{{"KSampler<br/>steps:20 cfg:8"}}
  end
  1 -->|MODEL| 5
  2 -->|CONDITIONING| 5
  3 -->|CONDITIONING| 5

Debug a failed workflow

> /comfy:debug

Reads the last execution history and logs, identifies the failing node, checks for missing models or node packs, and suggests a fix.

> /comfy:debug abc123-def456

Diagnose a specific execution by prompt ID.

Parameter sweep

> /comfy:batch a cat in a field, cfg:5-10:2, sampler:euler,dpmpp_2m

Generates a grid of images across all parameter combinations and presents a summary table with results.

Supported sweep parameters: cfg, steps, sampler, scheduler, seed, denoise, width, height.

Multi-step recipes

> /comfy:recipe hires-fix a dramatic fantasy landscape with castles

Runs two passes: txt2img at 512x768, then an img2img upscale to 1024x1536 that adds detail.

Available recipes:

| Recipe | Description |

|--------|-------------|

| portrait | Generate at 1024x1024, then 2x upscale to 2048x2048 |

| hires-fix | Low-res generation → img2img upscale with denoise 0.4-0.5 |

| style-transfer | Apply a style prompt to an existing image via img2img |

| product-shot | Product image with clean white background |

Convert workflow format

> /comfy:convert ~/workflows/my-ui-workflow.json

Converts between ComfyUI's UI format (nodes + links arrays) and API format (node IDs → {class_type, inputs}).

Install a custom node pack

> /comfy:install comfyui-impact-pack

Searches the registry, shows details, clones the repo to custom_nodes/, installs dependencies, and offers to restart ComfyUI.

Browse output gallery

> /comfy:gallery last 5
> /comfy:gallery today

Lists recent outputs with their embedded metadata: checkpoint, prompt, seed, steps, CFG, and sampler for each image.

Compare workflows

> /comfy:compare workflow-a.json vs workflow-b.json

Shows added/removed nodes, changed parameters (old → new values), and optional Mermaid diagrams for visual comparison.

Validate before running

> Validate this workflow before I run it

Checks for missing node types, broken connections, invalid output indices, and missing model files without executing the workflow.

Manage models

> What checkpoints do I have installed?
> Search HuggingFace for SDXL turbo models
> Download this model to my checkpoints folder

Manage VRAM

> Free my VRAM
> What embeddings do I have?

Extract workflow from an image

> Extract the workflow from this image: ~/outputs/ComfyUI_00042_.png

Reads the PNG metadata chunks to recover the exact workflow and prompt used to generate the image.

Explore custom nodes

> /comfy:node-skill comfyui-impact-pack

Generates a skill file documenting every node, its inputs/outputs, and usage patterns.

Process control

> Restart ComfyUI
> Stop ComfyUI
> Start ComfyUI back up

Configuration

The server auto-detects your ComfyUI installation and port. Override with environment variables if needed.

Where to put keys and overrides:

• Panel users. Use the API Keys card (▾ menu next to "connected"). Keys are stored in ~/.comfyui-mcp/panel-secrets.json and take effect without a restart.
• MCP-only setups (Claude Desktop, Claude Code, etc.). Use the env block of your MCP client config, as shown in the setup examples above.
• Real environment variables always take precedence over stored keys. For development, the server also loads ~/.comfyui-mcp/.env as an override file; regular installs shouldn't need it.

Deployment modes

comfyui-mcp operates in one of three modes, auto-selected from the environment:

| Mode | Trigger | Local FS / process tools? |

|------|---------|----------------------------|

| Local | default | yes |

| Remote | --comfyui-url / COMFYUI_URL points at a non-loopback host, or --force-remote is set | no — server skips COMFYUI_PATH auto-detection so stale local installs can't silently absorb uploads |

| Cloud | COMFYUI_API_KEY is set (targets Comfy Cloud) | no — HTTP primitives route via cloud.comfy.org over X-API-Key; WebSocket and local-only tools throw CLOUD_UNSUPPORTED |

Some setups (e.g. dstack driving ComfyUI on RunPod) port-forward

a remote ComfyUI back to localhost:8188, so the loopback check above reports a

local install that isn't one. Pass --force-remote (or set

COMFYUI_MCP_FORCE_REMOTE=1) alongside --comfyui-url/COMFYUI_URL to force

remote mode regardless of hostname:

npx -y comfyui-mcp@latest --comfyui-url http://localhost:8188 --force-remote

| Variable | Default | Description |

|----------|---------|-------------|

| COMFYUI_URL | | Full ComfyUI URL, e.g. https://comfy.example.com:8443 — overrides COMFYUI_HOST/PORT/SSL and skips auto-detection. A path prefix is preserved (e.g. https://host/comfyapi) for reverse-proxied instances. Non-loopback hosts opt into remote mode. |

| COMFYUI_MCP_FORCE_REMOTE | | Set to 1/true (or pass --force-remote) to force remote mode even when COMFYUI_URL/--comfyui-url resolves to a loopback host — for port-forwarded remote installs (e.g. dstack/RunPod) that are reachable at localhost. No effect without a COMFYUI_URL/--comfyui-url. |

| COMFYUI_HOST | 127.0.0.1 | ComfyUI server address |

| COMFYUI_PORT | (auto-detect) | ComfyUI server port (tries 8188, then 8000) |

| COMFYUI_PATH | (auto-detect) | Path to the ComfyUI data/base directory used for models, input/output and user state. In a conventional install this is also the checkout. Auto-detection suppressed in remote/cloud modes. |

| COMFYUI_CODE_PATH | COMFYUI_PATH | Optional path to the ComfyUI checkout (main.py, .venv, core git) when code and data live under different roots. Pip, venv, and core updates use this checkout (core updates prefer the checkout observed from the connected local server and use this as their fallback). Pack reads/writes — custom_nodes, comfy-cli --workspace, apply_manifest clone/checkout, workflow-lock pack commits — stay on the live --base-directory / COMFYUI_PATH data root (#1770). |

| COMFY_CLI_PATH | (auto-detect) | Path to the official comfy executable (comfy-cli >=1.11.1). Resolution also checks PATH and the selected workspace's .venv/venv. |

| COMFYUI_PYTHON | python | Python interpreter used by legacy git-clone dependency fallbacks. Point it at your ComfyUI venv's Python when needed. |

| COMFYUI_MCP_BRIDGE_HOST | 127.0.0.1 | Panel-bridge bind host. Set 0.0.0.0 (or a LAN IP) to run the orchestrator on a 24/7 server and connect panels from other machines — requires a token (below); the orchestrator prints a ready-to-paste ws://…/?token=… Bridge URL. |

| COMFYUI_MCP_BRIDGE_TOKEN | (generated when needed) | Shared secret gating every bridge connection (checked constant-time on the WS upgrade). Mandatory for a non-loopback COMFYUI_MCP_BRIDGE_HOST; pin it so the Bridge URL survives restarts. Never logged beyond the startup banner. |

| COMFYUI_MCP_DATA_DIR | ~/.comfyui-mcp | Base dir for per-instance data (the generations.db behind get_history action: "suggest") when there's no local COMFYUI_PATH (remote/cloud/undetected). Scoped per target under instances/<host_port>/. |

| COMFYUI_API_KEY | | Comfy Cloud API key. When set, cloud mode is active and the server talks to cloud.comfy.org. Never logged. |

| COMFYUI_CLOUD_URL | https://cloud.comfy.org | Override the Comfy Cloud endpoint (testing/staging). |

| COMFYUI_AUTH_TOKEN | | Generic auth token for a self-hosted ComfyUI behind a reverse proxy / API gateway (distinct from Comfy Cloud). When set, attached to every ComfyUI request. Never logged. |

| COMFYUI_AUTH_HEADER | Authorization | Header name for COMFYUI_AUTH_TOKEN (e.g. X-API-Key). |

| COMFYUI_AUTH_SCHEME | Bearer for Authorization, else none | Scheme prefix on the token value (e.g. Bearer, Token). |

| CIVITAI_API_TOKEN | | CivitAI API token for model downloads |

| HUGGINGFACE_TOKEN | | HuggingFace token for higher API rate limits |

| COMFYUI_DOWNLOAD_PROXY | | Optional HTTP/HTTPS proxy for model-download traffic only; standard proxy environment variables and static Windows WinINet settings are also honored |

| GITHUB_TOKEN | | GitHub token for skill generation (avoids rate limits) |

| REGISTRY_ACCESS_TOKEN | | Comfy Registry API key for node_pack (action: "publish") (env-only, never logged) |

| COMFYUI_DOWNLOAD_CACHE_DIR | ~/.comfyui-mcp/cache | Content-addressed model-download cache (dedup + concurrent coalescing) |

| COMFYUI_LRU_CACHE_SIZE_GB | 0 | Cap the download cache in GB; 0 disables LRU eviction |

| COMFYUI_STARTUP_CHECK_INTERVAL_S / …_MAX_TRIES | 1 / 60 | Readiness-probe interval + max tries when starting a local ComfyUI |

| COMFYUI_ALWAYS_RESTART | false | Auto-restart a crashed local ComfyUI (bounded by COMFYUI_RESTART_MAX_ATTEMPTS / COMFYUI_RESTART_WINDOW_S) |

| COMFYUI_MCP_STALL_S | 180 | Render-wedge watchdog: seconds a sampler step can re-emit the same progress before a STALL/BACKLOG note is prepended to the agent's next turn (clamped 15–3600s; live-tunable from the panel) |

| COMFYUI_MCP_INTERRUPT_S | 30 | Seconds queue (action:"cancel") waits for an interrupt to actually stop a job before escalating to /free and reporting it wedged |

| LOG_LEVEL | info | Logging verbosity: debug, info, warn, error |

Transports

The server speaks stdio by default, which is what Claude Code, Claude Desktop, and the MCP Inspector expect, so no flags are needed. For MCP gateways, remote/hosted setups, or fetch-based clients, opt into streamable-HTTP:

# stdio (default)
npx -y comfyui-mcp@latest

# streamable-HTTP on http://127.0.0.1:9100/mcp
npx -y comfyui-mcp@latest --http
npx -y comfyui-mcp@latest --http --host 0.0.0.0 --port 9100   # bind/port overrides

| Flag | Env | Default | Description |

|------|-----|---------|-------------|

| --http / --transport http | MCP_TRANSPORT=http | stdio | Serve streamable-HTTP at /mcp instead of stdio |

| --host <h> | MCP_HOST | 127.0.0.1 | HTTP bind host (use 0.0.0.0 to expose) |

| --port <n> | MCP_PORT | 9100 | HTTP port |

| --comfyui-url <url> | COMFYUI_URL | (auto-detect) | Target a specific (incl. remote) ComfyUI |

| --force-remote | COMFYUI_MCP_FORCE_REMOTE | false | Force remote mode for a loopback --comfyui-url (e.g. dstack/RunPod port-forwards to localhost) |

Other agents & local LLMs (Hermes, OpenClaw, Copilot CLI, Ollama)

comfyui-mcp supports non-Claude harnesses. One command

writes the server entry into the harness's own config (it merges into an existing

file rather than overwriting it):

npx -y comfyui-mcp setup hermes     # → ~/.hermes/config.yaml      (compact by default)
npx -y comfyui-mcp setup openclaw   # → ~/.openclaw/openclaw.json  (compact by default)
npx -y comfyui-mcp setup copilot    # → ~/.copilot/mcp-config.json (full by default)
# flags: --compact | --full, --comfyui-url <url>, --dry-run

Model requirements. Tool calling is a hard requirement; a model without tool calling

doesn't work. Thinking and vision are recommended. Without

thinking, multi-step tool chains degrade; without vision the agent can

generate but can't see its own outputs.

Compact tool mode. --compact /

COMFYUI_MCP_TOOL_MODE=compact registers 3 meta-tools (list_tools →

describe_tool → call_tool) instead of the direct tool set, pulling schemas

into context one tool at a time. That is the right trade for small local models. The

direct tool set is the default and is what frontier-model harnesses want.

**Run it locally

for free with our fine-tuned models.** ollama pull artokun/gemma4-comfyui-mcp:e4b

(also :e2b for ~2 GB VRAM, :12b for ~8 GB) is Gemma 4, QLoRA-trained on 1,055

server-verified trajectories over the full comfyui-mcp tool set, and is the

panel's Ollama default. Stock gemma4:*/qwen3:4b are also validated end-to-end

(npm run test:local-llm); gemma3 has no native tool calling and is

unsupported. The full guide, covering hosted-model guidance (DeepSeek/MiMo/GLM class),

per-harness setup, and troubleshooting, is

[Local LLMs & other agents](https://comfyui-mcp.artokun.io/docs/local-llms).

| Flag | Env | Default | Description |

|------|-----|---------|-------------|

| setup <agent> | | | Write the comfyui entry into hermes / openclaw / copilot config, then exit |

| --compact / --tool-mode compact | COMFYUI_MCP_TOOL_MODE=compact | full | Opt into the 3 meta-tools (list_tools / describe_tool / call_tool) instead of the direct tool set — recommended for small local models. --full is still accepted and is now a no-op |

Remote ComfyUI

Point the server at a ComfyUI running anywhere; no local install is required:

npx -y comfyui-mcp@latest --comfyui-url http://192.168.1.50:8188
npx -y comfyui-mcp@latest --http --comfyui-url https://comfy.example.com:8443

Behind a reverse proxy / API gateway (path prefix + auth header). For a

self-hosted ComfyUI exposed under a prefixed route with its own auth layer (this

is not Comfy Cloud, which is COMFYUI_API_KEY):

COMFYUI_URL=https://gateway.example.com/comfyapi \
COMFYUI_AUTH_TOKEN=your-token \
  npx -y comfyui-mcp@latest --http        # → Authorization: Bearer your-token, requests under /comfyapi

# custom header / scheme:
COMFYUI_URL=https://gateway.example.com/comfyapi \
COMFYUI_AUTH_HEADER=X-API-Key COMFYUI_AUTH_TOKEN=your-token \
  npx -y comfyui-mcp@latest --http        # → X-API-Key: your-token

Auto-detection

Port: Probes 8188 (CLI default) then 8000 (Desktop app default) via /system_stats.

Path: Checks common locations in order:

• ~/Documents/ComfyUI (macOS/Windows Desktop app data directory)
• ~/Library/Application Support/ComfyUI (macOS)
• ~/AppData/Local/Programs/ComfyUI/resources/ComfyUI (Windows Desktop app install)
• ~/AppData/Local/ComfyUI (Windows)
• ~/ComfyUI, ~/code/ComfyUI, ~/projects/ComfyUI, ~/src/ComfyUI
• /opt/ComfyUI, ~/.local/share/ComfyUI (Linux)
• Scans ~/Documents and ~/My Documents for any directory containing "ComfyUI"

Set COMFYUI_PATH to skip detection and use an explicit path.

How It Works

The server communicates with ComfyUI through its REST API and WebSocket interface:

• WebSocket. Enqueue workflows, receive real-time progress updates (step by step via the background monitor script), get execution results
• REST API. System stats, node definitions (/object_info), logs, history, queue management, workflow library, VRAM control (/free), embeddings
• File system. Read/write the models directory, detect installation paths, upload images, extract PNG metadata, browse outputs
• External APIs. HuggingFace (model search), ComfyUI Registry (custom node discovery), GitHub (skill generation), CivitAI (model downloads)

All communication with the MCP client (Claude Code) happens over stdio using the Model Context Protocol. Logs go to stderr so they never mix into the protocol stream.

Development

Prerequisites

• Node.js >= 22.0.0
• ComfyUI running locally

Setup

git clone https://github.com/artokun/comfyui-mcp.git
cd comfyui-mcp
npm install

Scripts

| Script | Description |

|--------|-------------|

| npm run dev | Run from source with tsx (hot reload) |

| npm run build | Compile TypeScript to dist/ |

| npm start | Run compiled output |

| npm test | Run unit tests (vitest) |

| npm run test:integration | Run integration tests (requires running ComfyUI) |

| npm run lint | Type-check without emitting |

| npm run generations:stats | Show local generation tracking statistics |

| npm run sync-agents | Sync Claude skills/commands/hooks to Google Antigravity, OpenCode, and other AI IDE formats that supports .agents files |

Local testing with Claude Code

Point Claude Code at your local build instead of the npm package:

{
  "mcpServers": {
    "comfyui": {
      "command": "node",
      "args": ["/path/to/comfyui-mcp/dist/index.js"],
      "env": {}
    }
  }
}

Or test the plugin directly:

claude --plugin-dir ./plugin

Project structure

scripts/
  generation-stats.mjs         # CLI: npm run generations:stats
src/
  index.ts                 # MCP server entry point (stdio transport)
  config.ts                # Auto-detection & environment config
  comfyui/
    client.ts              # ComfyUI WebSocket/HTTP client wrapper
    types.ts               # TypeScript interfaces
  services/
    workflow-executor.ts   # Execute workflows, handle images & errors
    workflow-composer.ts   # Templates (txt2img, img2img, upscale, inpaint)
    workflow-validator.ts  # Dry-run validation (missing nodes, models, connections)
    image-management.ts    # Upload images, extract PNG metadata, list outputs
    mermaid-converter.ts   # Workflow → Mermaid diagram
    workflow-converter.ts  # UI → API: de-virtualize Get/Set buses + Reroutes, expand subgraphs, resolve bypass (powers get_workflow's strip action)
    workflow-slicer.ts     # sliceWorkflow() — rgthree Fast-Groups-Bypass pipeline un-chunker (shared by the CLI + get_workflow's slice action)
    mermaid-parser.ts      # Mermaid diagram → Workflow
    model-resolver.ts      # HuggingFace search, local models, downloads
    generation-tracker.ts  # SQLite generation log, settings dedup, stats
    file-hasher.ts         # SHA256 hashing of .safetensors with cache
    workflow-settings-extractor.ts  # Extract settings from workflow JSON
    process-control.ts     # Stop, start, restart ComfyUI process
    registry-client.ts     # ComfyUI Registry API
    skill-generator.ts     # Generate node pack skill docs
  tools/                   # MCP tool registration (one file per group)
    workflow-execute.ts    # enqueue_workflow, get_system_stats
    workflow-visualize.ts  # visualize_workflow (render/render_hierarchical/mermaid/to_dsl/from_dsl)
    workflow-compose.ts    # create_workflow (create/modify/validate/node_info)
    workflow-validate.ts   # the validate action's body
    workflow-library.ts    # get_workflow (8 read actions), save_workflow (save/lock/verify_lock)
    image-management.ts    # get_image (7 read actions), upload_image (5 write actions)
    model-management.ts    # download_model, list_local_models (the two consolidated model tools)
    memory-management.ts   # clear_vram
    registry-search.ts     # search_custom_nodes (search/details)
    node-management.ts     # install_custom_node (install/update/fix/uninstall/enable/disable/list/…)
    node-pack.ts           # node_pack (scaffold/verify/publish/read/write/patch/git/…)
    generation-tracker.ts  # the get_history suggest/stats jobs
    diagnostics.ts         # get_system_stats (action:"logs"), get_history
    process-control.ts     # restart_comfyui (restart/start/stop)
    index.ts               # Registers all tool groups
  utils/
    errors.ts              # Custom error hierarchy with MCP integration
    logger.ts              # stderr-only logging (safe for stdio transport)
    image.ts               # Base64 encoding utilities
plugin/
  .claude-plugin/          # Plugin manifest
  .mcp.json                # MCP server config for plugin
  commands/                # Slash commands
    gen.md                 # /comfy:gen — image generation
    viz.md                 # /comfy:viz — workflow visualization
    node-skill.md          # /comfy:node-skill — skill generation
    debug.md               # /comfy:debug — failure diagnosis
    batch.md               # /comfy:batch — parameter sweeps
    convert.md             # /comfy:convert — format conversion
    install.md             # /comfy:install — node pack installation
    gallery.md             # /comfy:gallery — output browser
    compare.md             # /comfy:compare — workflow diff
    recipe.md              # /comfy:recipe — multi-step pipelines
  skills/                  # Knowledge bases
    comfyui-core/          # Workflow format, node types, pipeline patterns
    prompt-engineering/    # CLIP syntax, model-specific prompting
    troubleshooting/       # Error catalog with patterns and fixes
    model-compatibility/   # Compatibility matrix per model family
  agents/                  # Autonomous agents
    explorer.md            # Research custom node packs, generate skills
    debugger.md            # Diagnose workflow failures
    optimizer.md           # Analyze and optimize workflows
  hooks/                   # Pre/post tool-use hooks
    hooks.json             # Hook configuration
    vram-check.mjs         # VRAM watchdog before execution
    save-warning.mjs       # Save prompt before stop/restart
    job-complete-notify.mjs # Job completion notification via temp files
  scripts/                 # Background scripts
    monitor-progress.mjs   # Real-time WebSocket progress monitor
    launch-server.mjs      # MCP server launcher — global install if present, else npx with a
                           #   cold-start handshake rescue so a first run cannot time out (#1447)

Troubleshooting

"ComfyUI not detected on ports 8188, 8000"

Make sure ComfyUI is running. The Desktop app uses port 8000 by default; the CLI uses 8188. Set COMFYUI_PORT if you're using a custom port.

**"COMFYUI_PATH is not configured"**

The auto-detection couldn't find your ComfyUI data directory. Set COMFYUI_PATH to the directory containing your models/ folder (e.g., ~/Documents/ComfyUI).

"Multiple ComfyUI installations detected"

This is informational. The server uses the first one found. Set COMFYUI_PATH to pick a specific installation.

Model downloads fail

For HuggingFace gated models, set HUGGINGFACE_TOKEN. For CivitAI, set CIVITAI_API_TOKEN.

Workflow execution errors

Use /comfy:debug to diagnose failures. Or call get_history (action: "diagnose") / get_system_stats (action:"logs") directly to see the full error messages, including Python tracebacks from ComfyUI.

Out of memory (OOM)

Use clear_vram to free GPU memory before running large workflows. The VRAM watchdog hook warns you when less than 1GB is free before a workflow runs. See the troubleshooting skill for model-specific VRAM estimates.

Missing custom nodes

Use /comfy:install <pack> to install missing node packs from the registry. The debug command detects missing packs and suggests them.

Contributing

Contributions are welcome. See CONTRIBUTING.md for the dev setup, project

conventions, how to add an MCP tool, and the release process.

Quick version: fork → branch (feat/my-feature) → make changes (make sure npm run build and

npm test pass; run npm run docs:gen if you touched tools) → open a PR.

Maintainer

Built and maintained by **@artokun**, a regular contributor across the Comfy-Org ecosystem:

• **Comfy-Org/ComfyUI_frontend.** 10 merged PRs, mostly on the v2 graph renderer: subgraph rendering, promoted-widget plumbing, viewport persistence, with backports across cloud/1.41, cloud/1.42, core/1.41, and core/1.42.
• [Comfy-Org/ComfyUI](https://github.com/Comfy-Org/ComfyUI/pulls?q=is%3Apr+author%3Aartokun) (core). Crash fixes in the Python backend's video/audio save path (#12683, #12550).

If you're at Comfy-Org (or hiring around the ComfyUI ecosystem), I'd love to chat. Email me at [art.longbottom.jr@gmail.com](mailto:art.longbottom.jr@gmail.com).

License

MIT. See LICENSE for details.

Changelog

See CHANGELOG.md for the full, structured release history.