Claude Chat

This node ("Claude Chat" in the menu) drops Anthropic's Claude right into your graph as a chat-and-vision node. It's not a local model and it's not free: you paste an API key, every run makes a real API call to Anthropic, and your prompt (and any image you feed it) leaves your machine. That's the entire deal, so know it going in - this is the "bring an LLM in to write a prompt or look at an image" pattern, not another sampler.

So why reach for it? In the videoclipgenerator world it slots in nicely upstream of everything else: Claude writes the lyrics, the voiceover script, or the per-shot image prompts that VibeMusicEngine and MovingTitles then consume. Feed it a generated frame through the optional image input and it'll describe what's on screen - useful when you're trying to prompt a consistent style across a video's worth of shots. The output plumbing is unusually generous: response_text, plus input_tokens, cost_usd, stop_reason, and a conversation_id you can reuse to keep chatting.

The inputs that actually matter

• api_key - your Anthropic key. Leave it blank and it falls back to an ANTHROPIC_API_KEY environment variable; blank both and it throws.
• model - four choices baked in: claude-3-5-sonnet-latest, claude-3-opus, claude-3-sonnet, claude-3-haiku. The tooltip's cost framing is the point: Opus is the expensive brain, Haiku is the cheap fast one.
• prompt and system_prompt - your message and the persona/constraints it operates under. The default system prompt is "you are a helpful AI assistant in ComfyUI" which is fine to replace.
• max_tokens (default 1024) and temperature (default 0.7) - response length and randomness. Drop temperature toward 0 when you want structured, repeatable output.
• new_session and clear_history (no/yes/all) - control conversation state. conversation_id in the optional section lets you resume a specific thread later.
• summarize (default on) - auto-summarizes history as you approach the context window so a long dialogue doesn't hard-fail.

The optional image input takes an IMAGE tensor, with image_prompt being the specific question ("What can you see in this image?"). Note the tooltip's fine print: vision is only supported by Claude 3 models except Haiku.

What comes out

response_text is what you actually use. usage_info and input_tokens track spend, cost_usd is the nice-to-have that tells you what that run just cost you in dollars, history_messages reports how much context got carried, and model_info is a readout when you tick show_model_info.

Installing it

It ships inside the videoclipgenerator pack. ComfyUI Manager → search "videoclipgenerator", or:

cd ComfyUI/custom_nodes
git clone https://github.com/lazniak/videoclipgenerator
cd videoclipgenerator
pip install -r requirements.txt   # or: python install.py

Restart ComfyUI. The only pack dependency this node pulls in beyond the usual is the anthropic Python package.

Gotchas

The two things that bite people: forgetting the key (paste it in the widget or set ANTHROPIC_API_KEY in your environment) and forgetting you're paying per call - wire this into a batch workflow and watch cost_usd for a while before you trust it. Also, every node in this pack has a buy_me_a_coffee checkbox, and ticking it makes the node open a browser tab to the author's Buy Me a Coffee page. It's opt-in, it just surprises people. And the standard API-node warning applies: this is custom Python holding a credential and phoning home by design, so know what you installed - the LLMVISION incident should keep you honest about which packs you trust with a key.