Claude Fable 5

"Claude Fable 5" is the node you'll find yourself staring at, wondering what it is. Fair enough. It's a Claude-family model that KIE publishes an API page for, and this pack turns every published page into its own ComfyUI node. So there's a node. What you get is an LLM in your graph that takes a prompt and returns text - no weights on your disk, no VRAM, nothing local to babysit.

What it's for

Same job as every LLM node in this pack: short, structured text work that sits upstream of generation. Rewriting a rough idea into a prompt your checkpoint actually responds to. Turning scene notes into one clean paragraph for Wan. Summarising a mess of tags. It is not a text encoder - it doesn't condition your sampler, it writes the string a text-encode node then conditions on. If that distinction is fuzzy, the KB's llm-in-comfyui.md framing is worth ten minutes: the model inside your checkpoint and the model you drop into the graph are different animals.

Where this one sits in the Claude menu is a thing I can't honestly tell you. There's essentially no community chatter about it. KIE's own docs page is the authority, and the node's description links back to it.

How it works

The class is generated at startup from KIE's public API index, and for the Claude family it builds an Anthropic-shaped request: model, messages, stream, max_tokens, an optional system, tools, and a thinkingFlag. It POSTs that to KIE's /claude/v1/messages endpoint. Your API key never leaves api.kie.ai.

Two behaviours worth knowing before you debug anything:

• history_json is passed straight through as the messages array. Your prompt is then appended as the final {"role": "user"} turn. So history is prior turns only - if you paste the current question into both, the model sees it twice.
• thinking maps to KIE's thinkingFlag: true, not Anthropic's thinking: {type: "enabled", budget_tokens: N}. If you need the documented raw shape, that's what expert_override_json is for.

No IMAGE input on Claude nodes, deliberately - multimodal goes through history_json/expert_override_json, with the image uploaded first via KIE • Upload Image.

Inputs and outputs

Only prompt is required. The optional fields you'll actually touch: system_prompt (the persona/format instruction, and where you should put "respond with only the prompt, no preamble"), max_tokens (default 4096, capped at 131072 in the widget), and thinking. tools_json and expert_override_json are for later.

Outputs: text is your product - wire it into a CLIPTextEncode, a Wan/Seedance prompt socket, or a Suno prompt. raw_json and usage_json are for when something looks wrong. credits_consumed and credits_left are floats, and the node prints a "credits spent / credits left" receipt on itself after each run.

stream is a transport detail, not a live token feed: turn it on and the client reassembles the event stream and still hands you finished text when the call returns.

Install

Search ComfyUI-KIE-Nodes-Next in ComfyUI Manager. The README says Registry publication is prepared but not shipped, so if Manager comes up empty:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
cd ComfyUI-KIE-Nodes-Next
pip install -r requirements.txt   # requests, Pillow, numpy, PyYAML

Restart ComfyUI, then go to Settings → KIE.ai Nodes Next → Connection and paste your KIE key. It's validated before it replaces anything, and afterwards you get a masked fingerprint (••••••••••••A1B2) plus a CONNECTED line. KIE_API_KEY in the environment overrides the saved key.

Only dependency is Python 3.10+. Nothing downloads, nothing renders, no GPU.

Common issues

The node doesn't exist yet. On first launch the pack reads KIE's docs index and rebuilds the model library, then reloads the frontend once. Give it a minute before you go looking for missing nodes.

Every call costs credits. A local 8B prompt-enhancer is free per invocation; an API call is metered, every time. Treat each test run as money.

Your prompt goes to a third party. That's the mechanism, not a bug, and it means the endpoint's filters apply. For anything a hosted provider refuses, a local model is the answer - nothing in this pack bypasses a filter.

No key, no call. The error is explicit ("No KIE API key configured"). And since this is a small, unvetted pack - the corpus barely knows it exists - read the code before you trust it with a key you use elsewhere.