comfy_api_nodes/

Built-in nodes that call paid third-party AI APIs from inside ComfyUI workflows. ~40,000 lines, the second-largest package by line count after comfy/. The --disable-api-nodes flag turns the whole thing off.

Why this is its own package

Each provider needs:

• An HTTP client with auth (the user's api_key_comfy_org)
• Pydantic request/response schemas
• File upload (images, videos, audio) and download
• Polling for async jobs (most providers don't return synchronously)
• Error mapping into ComfyUI's exception model

Rather than scattering this across comfy_extras/, the team gave it a dedicated package with shared infrastructure under comfy_api_nodes/util/.

Layout

comfy_api_nodes/
├── __init__.py
├── nodes_<provider>.py     # One file per provider — the nodes themselves
├── apis/                   # Generated Pydantic models (regenerate via comfy_api/generate_api_stubs.py)
│   ├── __init__.py         # 6,155 lines, generated
│   └── schemas/            # Source schemas (input)
└── util/                   # Shared HTTP client, conversions, polling
    ├── __init__.py
    ├── _helpers.py
    ├── client.py           # 40,695 lines — the API client base
    ├── common_exceptions.py
    ├── conversions.py      # Image ↔ base64 ↔ URL, video reframe, audio
    ├── download_helpers.py # Robust downloads with retries
    ├── upload_helpers.py   # Pre-signed URL uploads, multipart
    ├── request_logger.py   # Per-request log lines for debugging
    └── validation_utils.py

Providers

There is roughly one nodes_<provider>.py file per integrated service. Sizes vary based on how many endpoints the provider exposes. The largest are video providers because video APIs typically expose many quality/length/style knobs.

Provider	File	Lines
Kling	comfy_api_nodes/nodes_kling.py	3,327
Wan (Comfy Org)	comfy_api_nodes/nodes_wan.py	2,222
ByteDance	comfy_api_nodes/nodes_bytedance.py	2,111
Vidu	comfy_api_nodes/nodes_vidu.py	1,728
Recraft	comfy_api_nodes/nodes_recraft.py	1,200+
Magnific	comfy_api_nodes/nodes_magnific.py	1,000+
Gemini	comfy_api_nodes/nodes_gemini.py	~900
Tripo	comfy_api_nodes/nodes_tripo.py	~800
ElevenLabs	comfy_api_nodes/nodes_elevenlabs.py	~800
Stability	comfy_api_nodes/nodes_stability.py	~700
Meshy	comfy_api_nodes/nodes_meshy.py	~700
Hunyuan3D (cloud)	comfy_api_nodes/nodes_hunyuan3d.py	~700
OpenAI	comfy_api_nodes/nodes_openai.py	~800
Grok	comfy_api_nodes/nodes_grok.py	~600
Ideogram	comfy_api_nodes/nodes_ideogram.py	~600
Veo (Google)	comfy_api_nodes/nodes_veo2.py	~500
Runway	comfy_api_nodes/nodes_runway.py	~400
BFL (Black Forest)	comfy_api_nodes/nodes_bfl.py	~500
Topaz	comfy_api_nodes/nodes_topaz.py	~400
Bria	comfy_api_nodes/nodes_bria.py	~250
Hitpaw	comfy_api_nodes/nodes_hitpaw.py	~280
Luma	comfy_api_nodes/nodes_luma.py	~500
LTXV	comfy_api_nodes/nodes_ltxv.py	~150
Minimax	comfy_api_nodes/nodes_minimax.py	~350
Moonvalley	comfy_api_nodes/nodes_moonvalley.py	~430
Pixverse	comfy_api_nodes/nodes_pixverse.py	~340
Quiver	comfy_api_nodes/nodes_quiver.py	~200
Reve	comfy_api_nodes/nodes_reve.py	~330
Rodin	comfy_api_nodes/nodes_rodin.py	~370
Sonilo	comfy_api_nodes/nodes_sonilo.py	~220
Sora	comfy_api_nodes/nodes_sora.py	~120
Wavespeed	comfy_api_nodes/nodes_wavespeed.py	~140

The full list is whatever files are in comfy_api_nodes/ at HEAD.

How a typical provider node works

sequenceDiagram
    participant Node
    participant Util as comfy_api_nodes/util/client.py
    participant API as Provider HTTPS API
    Node->>Util: build request from inputs
    Util->>Util: upload local images via upload_helpers
    Util->>API: POST /generate
    API-->>Util: { job_id }
    loop polling
        Util->>API: GET /jobs/{job_id}
        API-->>Util: { status, progress }
        Util-->>Node: set_progress
    end
    API-->>Util: { status: succeeded, result_url }
    Util->>Util: download_helpers.fetch(result_url)
    Util-->>Node: tensor / video / 3D mesh

Authentication uses the user's Comfy Org API key, threaded through the prompt as a hidden api_key_comfy_org input (see SENSITIVE_EXTRA_DATA_KEYS in execution.py). The client never logs the key; it does log full URL + method + status via request_logger.py.

CLI flag and feature flag

--disable-api-nodes turns this whole package off and prevents the frontend from communicating with the internet for API-node use. The flag is declared in comfy/cli_args.py and consulted in start_comfyui in main.py when initializing nodes.

API nodes also opt into the API_NODE = True flag on the ComfyNodeABC so the frontend can mark them visually.

Generated stubs

comfy_api_nodes/apis/__init__.py is 6,155 lines and entirely generated. It contains Pydantic models for every provider's request/response. Don't edit it by hand — edit comfy_api_nodes/apis/schemas/ and run python comfy_api/generate_api_stubs.py.

Integration points

• Loaded by nodes.init_extra_nodes(init_api_nodes=not args.disable_api_nodes) in server.py (called from main.py).
• Imports comfy_api.latest for the V3 schema and IO types.
• Imports aiohttp (already a runtime dependency).
• Each provider node typically imports its own Pydantic model from comfy_api_nodes/apis/.

Where to start adding a provider

1. Drop schemas under comfy_api_nodes/apis/schemas/.
2. Run python comfy_api/generate_api_stubs.py to regenerate comfy_api_nodes/apis/__init__.py.
3. Add comfy_api_nodes/nodes_<provider>.py modeled on a similar existing one (Kling, Veo, Stability are good references).
4. Use the helpers in comfy_api_nodes/util/: client.py for HTTPS, upload_helpers.py and download_helpers.py for blob I/O, conversions.py for image/video formats.
5. Make sure API_NODE = True is set on each node class.