Griptape Prompt Driver: Amazon SageMaker Jumpstart

This is the most niche prompt driver in the pack, and it's honest about it. Where the other drivers point at someone else's hosted API, this one talks to a model you deployed on Amazon SageMaker - most likely via Jumpstart, which is how AWS serves pre-trained open models like Llama on your own endpoints. If you have an endpoint already running and you're paying for it, this node makes that endpoint the brain of a Griptape agent. If you don't have a SageMaker endpoint, skip this one; standing one up just to use a node is a big bill looking for a problem.

The point of SageMaker here is control and compliance: the model runs in your AWS account, your data doesn't leave it, and you pick the instance size. That's a real selling point for anyone under data-handling constraints, and it's why this node exists.

How it works

You tell it the endpoint name of a deployed SageMaker endpoint (the default placeholder, jumpstart-dft-..., is exactly that - a placeholder), and it builds a Griptape AmazonSageMakerJumpstartPromptDriver that sends your agent's prompts to that endpoint and reads back the completion. Credentials come from the same AWS environment variables as the Bedrock driver.

The inputs that matter

• endpoint - the SageMaker endpoint you deployed. This is the one that must be right; there's no auto-discovery.
• model - a plain string, default meta-llama/Meta-Llama-3-8B-Instruct. It's informational metadata for the driver rather than something that spins up a model - the endpoint decides what actually runs. Keep it descriptive so a workflow reads clearly.
• max_tokens - default -1, meaning "use the model's default." Unlike the Bedrock driver, SageMaker doesn't refuse to answer if you leave this alone.
• The AWS fields again: aws_access_key_id_env_var, aws_secret_access_key_env_var, aws_default_region_env_var - env-var names, defaults AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_DEFAULT_REGION.

You also get the standard temperature, seed, max_attempts_on_fail, and use_native_tools. Output is a single DRIVER (PROMPT_DRIVER) to wire into agents and RAG modules.

Installing it

Same pack install as the rest: ComfyUI Manager → Griptape → ComfyUI-Griptape, or

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

restart, and set your AWS credentials in Settings → Griptape or as environment variables.

Common issues

Two things bite. First, the endpoint must be live - a stopped or deleted endpoint fails with a connection error, and Jumpstart endpoints don't stay hot forever. Second, make sure the model you deployed speaks the same chat format the driver expects; a raw base model with no instruction tuning will give you garbage completions even when the plumbing works. If the request succeeds but the text is nonsense, that's your model, not the node.

It's a narrow tool, but for the handful of people with a SageMaker endpoint and a ComfyUI habit, it's the missing link - no writing custom API glue to make an LLM node talk to AWS.