ComfyUI Core Knowledge

Workflow JSON Format (API Format)

ComfyUI workflows are JSON objects mapping string node IDs to node definitions:

{
  "1": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": { "ckpt_name": "sd_xl_base_1.0.safetensors" },
    "_meta": { "title": "Load Checkpoint" }
  },
  "2": {
    "class_type": "CLIPTextEncode",
    "inputs": { "text": "a cat", "clip": ["1", 1] },
    "_meta": { "title": "Positive Prompt" }
  }
}

Key Rules

• Node IDs are strings of integers ("1", "2", etc.)
• class_type is the exact Python class name of the node
• inputs contains both widget values (scalars) and connections (arrays)
• Connections use the format ["sourceNodeId", outputIndex], a 2-element array where:
  • the first element is the string node ID of the source node
  • the second element is the integer index into the source node's output list (0-based)
• _meta is optional and used for display titles only

Connection Examples

"model": ["1", 0]       // Connect to node 1's first output (MODEL)
"clip": ["1", 1]        // Connect to node 1's second output (CLIP)
"vae": ["1", 2]         // Connect to node 1's third output (VAE)
"positive": ["2", 0]    // Connect to node 2's first output (CONDITIONING)
"samples": ["5", 0]     // Connect to node 5's first output (LATENT)
"images": ["6", 0]      // Connect to node 6's first output (IMAGE)

Important: API Format vs Web UI Format

• API format (for execution/analysis) is { "1": { class_type, inputs }, "2": { ... } }. It is compact and used by enqueue_workflow, create_workflow (action:"validate"), create_workflow (action:"modify"), etc.
• Web UI format (for saving and frontend editing) is { "nodes": [...], "links": [...] }. It includes layout positions, sizes, groups, and visual metadata so ComfyUI's canvas can open and edit it
• Execution tools expect and return API format
• Save in Web UI format so saved workflows stay readable and editable in the ComfyUI frontend. A raw API-format save is not canvas-editable. It "exists" in the library but loads blank in the canvas, which strands users and tempts agents into creating yet another new workflow instead of reopening the old one. Because of this, save_workflow auto-converts API-format input to Web UI format with a generated layout. Prefer passing real Web UI format (from get_workflow(action="get", filename=…, format="ui")), since a generated layout loses the original node positions and groups
• get_workflow defaults to format="api" for analysis/execution; use format="ui" when loading a workflow to re-save or edit in the canvas
• Muted/bypassed nodes are preserved with _meta.mode: "muted". They are inactive but visible for understanding the workflow
• Get/Set virtual wire nodes are preserved with _meta.title and Constant key for tracing data flow

Workflow Library Tools

• get_workflow(action="analyze", filename=…) is the first call for understanding any saved workflow. It returns a structured text summary with sections, node IDs, key settings, virtual wires, and connection graph. No raw JSON, just what you need to reason about the workflow. Supports views: summary (default), overview (mermaid), detail (section mermaid), list, flat.
• get_workflow (action:"list") lists all saved workflows in ComfyUI's user library
• get_workflow(action="get", filename=…) loads raw workflow JSON. Only use it when you need the actual JSON for enqueue_workflow, create_workflow (action:"modify"), or save_workflow. Use action="analyze" instead for understanding. When the JSON is headed back to save_workflow, request format="ui" so the workflow stays editable in the frontend.
• save_workflow(action="save", filename=…, workflow=…) saves a workflow to the user library. Pass Web UI format ({ nodes, links }) so it keeps its real layout in ComfyUI's canvas. API-format graphs are accepted and auto-converted to Web UI format (with a generated layout) precisely because a raw API-format save is not canvas-editable; the frontend cannot open it. When re-saving an existing workflow, load it with get_workflow(action="get", filename=…, format="ui") and edit that, so positions and groups survive.

Data Types

ComfyUI nodes pass typed data through connections:

Type	Description	Common Source
MODEL	Diffusion model weights	CheckpointLoaderSimple (output 0)
CLIP	Text encoder	CheckpointLoaderSimple (output 1)
VAE	Variational autoencoder	CheckpointLoaderSimple (output 2)
CONDITIONING	Encoded text prompt	CLIPTextEncode (output 0)
LATENT	Latent space tensor	EmptyLatentImage, KSampler, VAEEncode
IMAGE	Pixel image tensor (BHWC)	VAEDecode, LoadImage, SaveImage
MASK	Single-channel mask	LoadImage (output 1)
UPSCALE_MODEL	Upscaling model	UpscaleModelLoader

Standard Pipeline Patterns

Text-to-Image (txt2img)

CheckpointLoaderSimple → MODEL, CLIP, VAE
  ├─ CLIP → CLIPTextEncode (positive) → CONDITIONING
  ├─ CLIP → CLIPTextEncode (negative) → CONDITIONING
  │
EmptyLatentImage → LATENT
  │
KSampler (model, positive, negative, latent_image) → LATENT
  │
VAEDecode (samples, vae) → IMAGE
  │
SaveImage (images)

Node IDs typically: 1=Checkpoint, 2=Positive, 3=Negative, 4=EmptyLatent, 5=KSampler, 6=VAEDecode, 7=SaveImage

Image-to-Image (img2img)

Same as txt2img but replace EmptyLatentImage with:

LoadImage → IMAGE
VAEEncode (pixels, vae) → LATENT → KSampler.latent_image

Set KSampler.denoise to 0.5 to 0.8 (lower = closer to input image).

Upscale

LoadImage → IMAGE
UpscaleModelLoader → UPSCALE_MODEL
ImageUpscaleWithModel (upscale_model, image) → IMAGE
SaveImage (images)

Inpaint

LoadImage (image) → IMAGE → VAEEncode → LATENT
LoadImage (mask) → MASK
SetLatentNoiseMask (samples, mask) → LATENT → KSampler.latent_image

MCP Tool Usage Guide

Quick Generation

1. create_workflow with template "txt2img" and your params
2. enqueue_workflow(action="enqueue") with the returned JSON. It returns prompt_id immediately
3. Poll queue (action:"status") with the prompt_id until done is true
4. Use get_image (action:"list_outputs") (limit 1) to find the generated image, then Read to display it

Inspect & Modify

• create_workflow (action:"node_info") queries what nodes are available and their schemas
• create_workflow (action:"modify") patches an existing workflow (set_input, add_node, remove_node, connect, insert_between)
• visualize_workflow shows a workflow as a mermaid diagram

Reverse Engineering

• visualize_workflow turns workflow JSON into a mermaid diagram
• visualize_workflow (action:"mermaid") turns a mermaid diagram into workflow JSON (uses /object_info for schema resolution)

Model Management

• list_local_models shows what's installed
• download_model action:"search" finds models on HuggingFace
• download_model downloads to ComfyUI's models directory

Never ask the user to manually download models. If a required model is missing, search for it and download it yourself:

1. Check list_local_models first
2. If missing, search HuggingFace via download_model action:"search" or CivitAI via their REST API
3. Use download_model to install it directly to the correct subfolder

CivitAI API (when the CIVITAI_API_TOKEN env var is available):

• Search: GET https://civitai.com/api/v1/models?query={query}&types=Checkpoint&sort=Most+Downloaded&limit=5
• Details: GET https://civitai.com/api/v1/models/{modelId}
• Download: GET https://civitai.com/api/download/models/{modelVersionId}?token={token}

CivitAI is preferred for fine-tuned models, community-rated checkpoints, and specialized LoRAs. HuggingFace is preferred for official/base models (SDXL, Flux, SD 1.5).

Custom Nodes

• search_custom_nodes searches the ComfyUI Registry (action: "search") or gets one pack's details (action: "details")
• list_packs (action: "generate_skill") auto-generates a skill file for a node pack

Workflow Execution

enqueue_workflow submits to ComfyUI's queue and returns prompt_id + queue position immediately. It does not block.

Background Progress Monitoring

After enqueuing one or more workflows, use a background Bash task to monitor progress silently:

# Single job
Bash(run_in_background: true):
node "${CLAUDE_PLUGIN_ROOT}/scripts/monitor-progress.mjs" <prompt_id>

# Multiple jobs (batch)
Bash(run_in_background: true):
node "${CLAUDE_PLUGIN_ROOT}/scripts/monitor-progress.mjs" <id1> <id2> <id3>

The script connects to ComfyUI's WebSocket and reports:

• Step-by-step progress (e.g., KSampler step 12/20 (60%))
• Success with output filenames and timing
• Errors with node details and messages

The standard generation pattern:

1. create_workflow or build workflow JSON + enqueue_workflow(action="enqueue") (repeat for batch)
2. Start background monitor with all prompt_ids
3. Continue conversation. Results appear when jobs finish
4. Use get_image (action:"list_outputs") or Read to display the generated images

Do not poll queue (action:"status") in a loop. The background monitor replaces polling entirely.

If the monitor script is unavailable, fall back to queue (action:"status") and poll until done is true.

Queue Management

One tool, queue, driven by its action parameter:

• queue (action:"list") shows running/pending job counts and prompt_ids
• queue (action:"status") checks if a specific prompt_id is running, pending, or done
• queue (action:"cancel") interrupts a running job (pass optional prompt_id to target a specific one)
• queue (action:"cancel_queued") removes a specific pending job from the queue by prompt_id
• queue (action:"clear") removes all pending jobs (does not stop the currently running job)

When to use queue tools:

• To check status, use queue (action:"status") for a quick boolean check (prefer the background monitor for ongoing tracking)
• To abort, queue (action:"cancel") stops what's running now and queue (action:"cancel_queued") removes a pending one
• To start fresh, queue (action:"clear") then optionally queue (action:"cancel")

Monitoring & Recovery

• get_system_stats reports GPU, VRAM, Python version, OS details
• queue (action:"list") shows running/pending jobs (also listed above under Queue Management)

When ComfyUI is unresponsive or crashed:

1. Try get_system_stats. If it fails, ComfyUI is down
2. Use restart_comfyui with action: "restart" (preserves launch args from a prior action: "stop")
3. If restart fails (no saved process info), use restart_comfyui with action: "start" or ask the user to start it manually
4. After ComfyUI is back, re-enqueue any failed/lost workflows

When a job appears hung (monitor shows [STALL]):

1. Check get_system_stats and look at VRAM usage (OOM causes hangs)
2. Try queue (action:"cancel") to interrupt the stuck job
3. If cancel fails, use restart_comfyui to force-restart
4. Use clear_vram after restart to free GPU memory before retrying

KSampler Parameters

Parameter	Type	Common Values
seed	int	Random (0 to 2^48). Omit to auto-randomize.
steps	int	20 (standard), 4-8 (turbo/lightning models)
cfg	float	7-8 (SD 1.5/SDXL), 1.0 (Flux), 3.5 (turbo)
sampler_name	string	"euler", "euler_ancestral", "dpmpp_2m", "dpmpp_sde"
scheduler	string	"normal", "karras", "sgm_uniform"
denoise	float	1.0 (txt2img), 0.5-0.8 (img2img), 0.75-0.9 (inpaint)

Mermaid Visualization Conventions

The visualize_workflow tool produces mermaid flowcharts with:

• Subgraphs grouping nodes by category: loading, conditioning, sampling, image, output
• Edge labels showing data types: -->|MODEL|, -->|CLIP|, -->|LATENT|, etc.
• Node labels showing class_type and optionally widget values
• Direction LR (left-to-right) by default, TB (top-to-bottom) for large workflows

The visualize_workflow (action:"mermaid") tool parses mermaid back into workflow JSON, using connection type labels to resolve the correct input/output slots via /object_info schemas.

Common Mistakes to Avoid

1. Wrong connection format. Use ["1", 0] not [1, 0]; node IDs are strings
2. Web UI format. Don't pass { nodes: [], links: [] }; use API format
3. Missing VAE. CheckpointLoaderSimple has 3 outputs: MODEL(0), CLIP(1), VAE(2)
4. Wrong output index. Check the node's output list order via create_workflow (action:"node_info")
5. Seed handling. enqueue_workflow randomizes seeds by default unless disable_random_seed: true

Sources

• Official: ComfyUI workflow/API conventions from https://github.com/comfyanonymous/ComfyUI and https://docs.comfy.org
• Empirical: MCP tool recipes and KSampler default tables are product/empirical notes, not a vendor prompting guide.