☁️Aisuite Loader

Most of comfyui_LLM_party's model loading goes through the "OpenAI-compatible base_url + api_key" pattern, which covers a surprising amount of the LLM world - OpenAI itself, Ollama, DeepSeek, Qwen, GLM, Moonshot, and any relay API that speaks the same dialect. It does not cover Anthropic's Claude, AWS Bedrock, Google Vertex AI, or Hugging Face's hosted inference, because those don't expose an OpenAI-shaped endpoint. Aisuite Loader is the node that fills that gap, by wrapping aisuite - a small library whose entire job is giving a handful of unrelated LLM APIs one consistent call signature.

How it works

You pick a provider, give it a model_name, and the node loads a model handle through aisuite's unified interface. Everything downstream in your LLM Party graph - the agent, the chat loop, the tool-calling - talks to this model exactly the same way it would talk to an OpenAI-format model, without caring that the actual request underneath is hitting a completely different SDK and auth scheme. That's the whole point of aisuite: one function call, six providers, instead of learning six client libraries.

The inputs and outputs that matter

• provider - openai, anthropic, aws, azure, vertex, or huggingface. This decides which of the optional credential fields actually get used.
• model_name (default gpt-4o-mini) - the model id as that provider names it, e.g. a Claude model string for anthropic, a Bedrock model id for aws.
• The optional fields are provider-specific, and you only fill in the set that matches your choice: base_url / api_key for openai/anthropic/azure; aws_access_key_id / aws_secret_access_key / aws_region_name for AWS; google_project_id / google_region / google_application_credentials for Vertex; hf_api_token for Hugging Face.
• One output: model - a loaded model handle you wire into whatever LLM node in the party actually runs the conversation.

If you only ever call OpenAI-compatible endpoints, you probably don't need this node - the plain LLM API loader is simpler. Reach for aisuite_loader specifically when you want Claude, Bedrock, Vertex, or an HF-hosted model in the same graph as everything else.

How to install it

• ComfyUI Manager - search "comfyui_LLM_party", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/heshengtao/comfyui_LLM_party, then restart.

Then, from inside the comfyui_LLM_party folder, run pip install -r requirements.txt in your ComfyUI Python environment (on a portable Windows install that's python_embeded\python.exe -m pip install -r requirements.txt). The pack bundles aisuite as one of its dependencies, so you don't need to install it separately - it comes along with the rest of the party's requirements.

Common issues & troubleshooting

Wrong fields for the provider. aisuite routes credentials by provider, so filling in api_key while provider is set to aws does nothing - the node needs the AWS-specific triple (aws_access_key_id, aws_secret_access_key, aws_region_name) instead. Match the credential fields to whatever you picked in provider, not to habit.

Vertex/AWS calls fail silently or with an opaque auth error. These two providers authenticate against your cloud account's IAM, not a simple bearer token, so the usual culprit is a service account or credentials file that doesn't have the right permissions scoped - that's a cloud-console problem, not a ComfyUI one, and worth ruling out before assuming the node is broken.

The pack itself won't load. This is a real, current issue reported by users installing via Manager and then running requirements manually - the pack fails to import with no clear list of supported CUDA/PyTorch versions to check against. If none of comfyui_LLM_party's nodes show up, that's almost always a base-install/dependency mismatch, not anything specific to aisuite_loader.

Model name doesn't match what the provider expects. Because model_name is a free-text field, a typo or an outdated model id (a deprecated Claude version string, say) produces a provider-side "model not found" rather than a ComfyUI error - check the id against the provider's current model list if calls are failing outright.