Griptape Agent Config: HuggingFace [DEPRECATED]

Another node you'll meet inside an old workflow rather than by adding it fresh. This is the deprecated Hugging Face structure config - the pre-"Drivers" version that only wires up a prompt driver against Hugging Face serverless inference. It still loads and runs, but it's been superseded by "Griptape Agent Config: HuggingFace Drivers," which adds an embedding driver so agents get memory and RAG without extra nodes.

What it is

A single config node that points a Griptape agent's prompt driver at an HF serverless inference model. The inputs are the old-style set: prompt_model (a free-text model ID, default HuggingFaceH4/zephyr-7b-beta), temperature, seed, max_attempts_on_fail, env (for environment variables), use_native_tools, max_tokens, and api_token_env_var (default HUGGINGFACE_HUB_ACCESS_TOKEN). It outputs a single CONFIG.

The same serverless caveats from the README apply here - only models supported by HF's serverless inference, and only models under 10GB. Type something outside that and you'll get inference errors.

Why it's deprecated

Same story as the other [DEPRECATED] configs in the pack: the "Drivers" family replaced these single-purpose configs with modular per-provider ones. The new HuggingFace Drivers node gives you both a prompt driver and an embedding driver (with its own model/tokenizer fields) in one box - meaning real agent memory out of the box. This old node has no embedding driver at all, so a memory-equipped agent had to be assembled from separate parts. It's kept around purely so legacy workflows don't break.

Installing it

It ships in the one pack:

cd ComfyUI/custom_nodes
git clone https://github.com/griptape-ai/ComfyUI-Griptape

or via ComfyUI Manager → search "Griptape". Restart, set your HF token (Settings → Griptape or HUGGINGFACE_HUB_ACCESS_TOKEN env var), and old workflows referencing it will load.

Should you use it?

Only if you're preserving an existing workflow. The migration is cheap: the new HuggingFace Drivers node drops into the same CONFIG socket, defaults to the same zephyr-7b-beta model, and adds embeddings for free. Building fresh, the deprecated node is strictly less useful - the only thing it saves you is a couple of optional fields you wouldn't touch anyway.

Common issues

• prompt_model typos - free-text input, no validation. A bad model ID fails at inference time with an HF error rather than a friendly message.
• No agent memory - expected with the old design; add embedding/vector store nodes manually or switch to the Drivers config.
• Serverless limits - >10GB or non-serverless models error out. Stick to small, popular models.
• Token not set - the standard [WARNING]: HUGGINGFACE_HUB_ACCESS_TOKEN is not set console message; verify the settings key and env var name.