Griptape Agent Config: Cohere Drivers

The Cohere Drivers config bundles everything a Cohere-backed Griptape agent needs into one node: a prompt driver running Cohere's command models, an embedding driver for RAG memory, and a local vector store wired to those embeddings, all behind a single COHERE_API_KEY. It's the "Cohere edition" of the pack's all-in-one config family - the equivalent of OpenAI's or Anthropic's, but for the Command/embedding family.

Why Cohere? Two honest answers. If you want RAG memory and retrieval quality, Cohere's embed-english-v3.0/embed-multilingual-v3.0 models are a legitimate reason to pick this vendor over OpenAI - the multilingual embeddings are genuinely good, and the input-type-aware design (search_query vs search_document) is a real retrieval-quality lever. And the command models have a pragmatic follow-through: they support native tool calling, which is exactly what an agent needs to actually use the tools you wire in. This config is the fastest way to get all of that working at once.

How it works

The node constructs a CohereDriversConfig from Griptape's framework: a CoherePromptDriver for reasoning, a CohereEmbeddingDriver for turning text into vectors, and a LocalVectorStoreDriver built on those embeddings for storing and searching them. Because embedding + local vector store come together in one node, an agent built on this config has working memory/RAG out of the box - you just need a task or RAG module to put it to use. The API key is shared across the drivers from the single cohere_api_key_env_var.

The inputs that matter

• model - the Command model for the prompt driver: command-r-plus (default, the big one), command-r, command, command-light, plus -nightly variants. Start with command-r for a good speed/quality balance; command-r-plus when you want the best reasoning.
• temperature - default 0.1.
• min_p / top_k - Cohere-native sampling knobs; top_k lower = more deterministic.
• embedding_model - embed-english-v3.0 (default) or embed-multilingual-v3.0; multilingual for non-English corpora.
• input_types - search_query / search_document / classification; match it to how the embeddings get used.
• cohere_api_key_env_var - env var name, default COHERE_API_KEY.

Output is a single CONFIG for agent creation.

A gotcha hiding in the description

Look closely at the node's description in ComfyUI: "OpenAI Structure Config. Use OpenAI's models for prompt, embedding, image generation, and image query." That's a copy-paste bug - the source file literally reuses the OpenAI description. It's Cohere, and there is no image generation or image query here. Don't let the text convince you the node does something it doesn't; the only Cohere inputs are the command model, the embedding model, and the key.

Install and key

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

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

Key from dashboard.cohere.com/api-keys, set as COHERE_API_KEY in Settings → Griptape or your environment. Standard pack caveats: griptape[all] drags in torch (reinstall from cu121 index on Nvidia if needed).

Common gotchas

Beyond the description bug: remember Cohere's command models are API-credit billed, and the embedding driver only matters if you actually run vector-store or RAG tasks - otherwise the embeddings half of this config is idle. And as with every driver-based config, if you set input_types to search_query for everything including stored documents, retrieval quality drops; Cohere v3 wants query-typed and document-typed inputs to differ.