Griptape Agent Config: OpenAI Compatible Drivers

This is the config node that quietly does the most work in the whole pack, because it's the one you point at anything that speaks OpenAI's API dialect. The OpenAI Drivers config is hard-wired to api.openai.com. This one isn't. Change base_url and suddenly your Griptape agent is talking to a local LM Studio server, a vLLM instance, a llama.cpp server, Groq, Together, or some random proxy that's OpenAI-compatible. In 2026 that's nearly every LLM server that matters, which makes this the config to reach for when you don't want to be locked to one provider.

How it works

It builds the same style of DriversConfig as the flagship OpenAI node - prompt, image generation, embedding, text-to-speech, and audio transcription drivers - but every driver talks to the base_url you supply instead of a fixed endpoint. api_key_env_var still defaults to OPENAI_API_KEY, so the key handling matches everything else in the pack: put the key in ComfyUI Settings → Griptape, or in the environment, and the node resolves the name you typed.

That one base_url field is the whole trick. Set it to https://api.openai.com/v1 and you've basically rebuilt the OpenAI Drivers config. Set it to http://127.0.0.1:1234/v1 with LM Studio running and you're fully local.

The inputs that matter

• base_url (default https://api.openai.com/v1) - the server you're pointing at. This is the input you'll actually change.
• model (default gpt-4o) - the prompt model name the target server recognizes.
• api_key_env_var (default OPENAI_API_KEY) - env var / settings key name. Most local servers accept any junk key or none.
• image_generation_model + size, text_to_speech_model + voice, audio_transcription_model, and embedding_model - the per-capability drivers, all optional.

Plus the shared set: temperature, seed, max_tokens, min_p, response_format, use_native_tools, max_attempts_on_fail. Output is a single CONFIG.

Installing it

Pack install is the usual: Manager → search "Griptape", or:

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

Restart, done. griptape[all] and python-dotenv come along automatically. One caveat the README flags: griptape drags in a torch build that can conflict with ComfyUI's - if things break right after install, that's the first suspect.

Common issues

• Connection refused with a local server - the server isn't running, or the port in base_url is wrong. LM Studio's default is http://127.0.0.1:1234/v1; Ollama's OpenAI endpoint is http://127.0.0.1:11434/v1.
• Model not found - the target server doesn't know the name you typed. On LM Studio the name has to match what's loaded; on vLLM it has to match a served model.
• Native tools fail on local models - small local models are often flaky at LLM-native tool calling. Flip use_native_tools off and Griptape's own tool engine takes over.

The trap to avoid: don't reach for a provider-specific config when "compatible" will do. This one node covers LM Studio, vLLM, and a dozen hosted services, so it's often the better default than committing to a single vendor's box.