Claude Code Plugin

comfyui-mcp is a full Claude Code plugin: 42 AI skills, 11 slash commands, 4 autonomous agents, and 3 hooks on top of the 38 MCP tools — expert ComfyUI knowledge that grows with every release.

comfyui-mcp isn’t just an MCP server — it ships as a complete Claude Code plugin. Installing the plugin gives Claude expert, model-specific knowledge about ComfyUI so it picks the right sampler, CFG, resolution, and model files without trial and error.

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

42 AI skills — and growing

Skills are curated knowledge documents Claude loads on demand. Each model family gets generation parameters, node graphs, curated model download URLs, and failure-mode guidance:

Skill	What Claude learns
flux-txt2img	Flux.1 Dev/Schnell + Flux 2 Klein — BF16 vs FP8 pitfalls, dual-CLIP wiring, VAE selection
wan-t2v-video / wan-flf-video	WAN 2.x text-to-video and first-last-frame workflows
wan-scail-replacement	SCAIL-2 in-video character replacement — reference-framing→scale rule, tuning and compositing pitfalls
ltxv2-video	LTX-2.3 (and LTX-2 19B) — GGUF UNet, distilled model, camera-control LoRAs, two-stage upscaling, kornia fix
minimax-h3-video	MiniMax H3 (Hailuo) local T2V/I2V/R2V — Comfy-Org INT8, turbo LoRAs, 15 s stereo clips; cites MiniMax’s official prompting guides
qwen-txt2img / qwen-image-edit	Qwen-Image generation and instruction-based editing
z-image-txt2img	Z-Image Turbo workflows
ideogram-ultra	Ideogram 4 — open-weight text-to-image, area prompting, logos / posters / readable text
ernie-image	ERNIE-Image — fast text-to-image, precise multilingual text rendering, runs on under 8 GB VRAM
anima-base	ANIMA 1.0 — ~2B anime/illustration model, Danbooru tags + natural language, anime inpainting, under 6 GB VRAM
anima-lora-trainer	Train custom anime LoRAs — kohya sd-scripts Gradio trainer, under 6 GB VRAM
ai-toolkit-trainer	Train custom WAN 2.2 + Z-Image LoRAs (people / styles / motion) — ostris AI-Toolkit, local or RunPod
installer-packs	Use, build, and derive one-command installer packs — and invite users to contribute new packs upstream
panel-node-pack-sync	Keep the sidebar panel node-pack in step with the orchestrator after an update — version-mismatch detection, a version pin that is warned about rather than overridden, and a sync whose result is re-read from disk
model-compatibility	Which VAE/CLIP/text-encoder pairs with which architecture — the #1 source of broken workflows
model-registry	Curated download URLs + target dirs for every model the skills reference — grows each release
civitai	Native Civitai discovery → local download/wire/generate loop via built-in download_model action:"search_civitai" (no key, no extra server); the official Civitai MCP is an optional community-surface add-on, not bundled
comfyui-core	Queue, history, and API semantics
comfyui-node-registry	Authoring + publishing custom node packs to the Comfy Registry
comfyui-frontend-extensions	v1/v2 frontend extension authoring (sidebar tabs, widgets)
rgthree	Configuring rgthree-comfy — Fast Groups Bypasser/Muter group toggles (frontend-only nodes absent from /object_info, set via node PROPERTIES not widgets), Power Lora Loader stacking, Context buses
lora-manager	ComfyUI-LoRA-Manager from the panel — add LoRA Text Loader (LoraManager) (STRING lora_syntax); Lora Loader (LoraManager) and other AUTOCOMPLETE_TEXT_* nodes wait 5s then refuse even when the pack is healthy
director	Multi-shot scene direction for video pipelines
prompt-engineering	Architecture-specific prompting (natural language vs tags)
troubleshooting	OOM, missing nodes, version drift — diagnosis trees
comfyui-launch-flags	The VRAM/attention/cache/perf launch-flag matrix — --reserve-vram, --novram+--cache-none, --use-sage-attention vs --use-pytorch-cross-attention (Z-Image), plus Blackwell/RTX 5000 accel-stack notes

New skills land with each release — generated from popular registry packs via the built-in list_packs tool (action: "generate_skill"), then human-curated.

Try the knowledge before wiring anything up — every skill is a plain markdown file, useful on its own. The most broadly applicable one is prompt-engineering — CLIP’s 77-token limit and BREAK chunking, weight syntax, and per-architecture strategy (why Flux wants natural language and no negative prompt while SD1.5 wants tags). Read it in the browser, or paste it into any agent’s context as a system prompt — no ComfyUI, MCP, or install required. When you do connect the full plugin, agents load the same files automatically (list_packs with action: "skill_read").

Civitai search — built in

download_model action:"search_civitai" searches Civitai natively (public REST API, no key, no extra server): keyword + types + base_models filters (“a Flux LoRA”), SFW-only by default, and every hit returns the model_version_id that action:"download_civitai" takes directly — plus the model’s trigger words for the prompt. Because it’s a first-class tool it works on every backend, including small local models behind the compact router. The civitai skill teaches the full discovery → download → generate handoff.

Optional API key. Search needs none. CIVITAI_API_TOKEN unlocks gated/early-access downloads and gated search results — one variable powers both search and download.

Optional: the official Civitai MCP. For the community surface beyond search→install (browsing example images + their generation params, posting, collections), pair Civitai’s official remote server — no longer auto-bundled, add it once:

claude mcp add --transport http civitai https://mcp.civitai.com/mcp \
  --header "Authorization: Bearer YOUR_CIVITAI_API_KEY"

11 slash commands

/comfy:gen (generate from a prompt), /comfy:debug (diagnose a failed workflow), /comfy:install (install node packs), /comfy:viz (workflow → Mermaid), /comfy:batch, /comfy:compare, /comfy:convert, /comfy:gallery, /comfy:recipe, /comfy:director, /comfy:node-skill.

4 autonomous agents

• comfy-explorer — deep-dives a node pack’s source to document it
• comfy-researcher — problem statement → ranked pack recommendations
• comfy-debugger — root-causes failed workflows from history + logs
• comfy-optimizer — VRAM and speed tuning for a given GPU

The sidebar panel — no API keys

The sidebar panel ships as its own pack, on the Comfy Registry and in ComfyUI-Manager as comfyui-agent-panel — search for it there and pick Latest, not Nightly. Full guide: Sidebar Panel.

The comfyui-mcp-panel sidebar runs an autonomous agent on your Claude subscription (no per-token API billing). Install the pack, open the Agent tab, and click Connect — it starts the panel orchestrator on demand; ask it for an image, a workflow, or a change and it works against your ComfyUI.

Hooks

Three hooks keep generation sessions tight: VRAM pre-flight checks before enqueueing a workflow, job-completion notifications, and save warnings before destructive operations.