Griptape Agent Config: Amazon Bedrock Drivers

The Amazon Bedrock Drivers config is the modern, all-in-one way to run a Griptape agent on AWS Bedrock. Where the deprecated Bedrock config only handled the prompt driver, this one builds a full stack: a prompt driver (Claude or Titan), a Titan embedding driver, a Titan image-generation driver, and a local vector store wired to the embeddings - all behind your AWS credentials. It's the node to use if your organization lives in AWS and you'd rather not route agent traffic through per-vendor SaaS keys.

The appeal is real if you're already an AWS shop: one credential set, billing consolidated on your AWS account, and Bedrock gives you Claude-class models plus Titan's image generation in one place. The catch is that this is the fiddliest of the pack's configs, because AWS has more moving parts than "paste an API key." The node tries to make that manageable by exposing the env var names and region explicitly.

How it works

The node constructs an AmazonBedrockDriversConfig with three drivers built from your inputs: AmazonBedrockPromptDriver for reasoning, AmazonBedrockTitanEmbeddingDriver for vectors, and AmazonBedrockImageGenerationDriver (the Titan image model) for image tasks, plus a LocalVectorStoreDriver on those Titan embeddings. Credentials come from the three env var names you provide - AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_DEFAULT_REGION by default.

The inputs that matter

• model - the prompt model: anthropic.claude-3-5-sonnet-20240620-v1:0 (default), other claude-3 IDs, or amazon.titan-text-*.
• max_tokens - default 100, and the tooltip is worth reading: "Amazon Bedrock tends to fail if this isn't given." This is the fix the deprecated node never got. Leave it at 100 or raise it for long outputs, but don't zero it.
• embedding_model - Titan embedding model (default amazon.titan-text-premier-v1:0).
• image_generation_model - default amazon.titan-image-generator-v1, with a size dropdown (512x512 / 1024x1024).
• temperature, min_p, top_k - the sampling dials, with Bedrock's own semantics.
• The three aws_*_env_var fields - env var names, never literal keys.

Output is a single CONFIG for agent creation.

Install and credentials

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

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

AWS access keys: console → your username → Security Credentials → Create Access Key (the README walks through it). Set the three vars in Settings → Griptape or your environment. griptape[all] installs torch; on Nvidia reinstall from the cu121 index if ComfyUI breaks.

Common gotchas

Bedrock failures are usually one of three things: the region doesn't host the model you picked (check Bedrock's model availability per region), the IAM role lacks Bedrock permissions, or max_tokens isn't set - which this node now handles by defaulting to 100. Titan image generation is also noticeably more limited than dall-e: the size dropdown tops out at 1024x1024 and there's no portrait option. And remember this is a config node - it emits CONFIG, so pair it with Griptape Create: Agent from Config, not with individual driver wiring.