Griptape Agent Config: Anthropic Drivers

The Anthropic Drivers config is the all-in-one way to build a Griptape agent whose brain is a Claude model. It replaces the old Anthropic [DEPRECATED] structure config with the modern driver-based approach: one node that produces a complete CONFIG - prompt driver, sampling parameters, key handling - ready to hand to agent creation. If you're an Anthropic user, this is the node you'll live in.

The big selling point is image understanding. The node's description calls it out: "Use Anthropic's models for prompt and image query." Claude's vision models are excellent at describing images, which in a ComfyUI context is gold - an agent that can look at a reference image and write you a prompt for it is one of the highest-value workflows this whole pack enables. Griptape Run: Image Description tasks powered by Claude are a big reason people install this pack at all.

How it works

The node wraps Griptape's AnthropicDriversConfig and builds an AnthropicPromptDriver from your settings, then registers it as the active drivers config. Because it's a config node, you don't wire individual drivers - everything is configured in place. The model dropdown is populated from Griptape's Claude model list (claude-3-5-sonnet family, claude-3-opus/sonnet/haiku, and later snapshots; the default is claude-3-5-sonnet-latest).

The inputs that matter

• model - the Claude model for the prompt driver. The default is fine for most work; opus when you want the best reasoning, haiku when you want speed on cheap tasks.
• temperature - default 0.1. Low for description/summarization, higher for creative prompt-writing.
• top_k - default 250, controls how many tokens are considered per step. The tooltip is the useful bit: higher lets the model avoid fixating on its top picks.
• top_p - default 0.999, the cumulative-probability cutoff. These two are Claude's native sampling knobs, exposed properly here.
• max_tokens - -1 lets the tokenizer default decide; raise it if long outputs get cut.
• api_key_env_var - env var name for your Anthropic key, default ANTHROPIC_API_KEY.

Output is a single CONFIG, feeding Griptape Create: Agent from Config.

Install and key

Same pack, same install - ComfyUI Manager (search "Griptape") or:

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

Set ANTHROPIC_API_KEY in Settings → Griptape or your environment; get the key from console.anthropic.com/settings/keys. The usual pack caveats apply: griptape[all] brings its own torch (reinstall from the cu121 index on Nvidia if ComfyUI breaks), and the env-var field takes a name, never a literal key.

Common gotchas

The big one is model-name drift. Anthropic renames and sunsets models fast, and the dropdown can lag reality - if the node's list doesn't have the model you want, update Griptape (python -m pip install griptape -U), since the list comes from the framework, not the node. Also remember: Anthropic accounts are API-credit billing, and image-query tasks chew through tokens quickly because the image gets sent to the API each time. If your agent's vision tasks start costing noticeable money, that's working as designed - consider the haiku-tier model for high-volume image description.