NVIDIA NeMo Agent Toolkit (nvidia-nat)

raw JSON →
1.6.0 verified Fri May 01 auth: no python

NVIDIA NeMo Agent Toolkit (nvidia-nat) is a Python library for building, training, and deploying conversational AI agents, including large language model (LLM) agents with tool use, memory, and multi-turn dialogue. Version 1.6.0 requires Python <3.14, >=3.11. Releases occur approximately monthly.

pip install nvidia-nat
error ImportError: cannot import name 'AgentConfig' from 'nvidia_nat'
cause AgentConfig was moved to nvidia_nat.config in version 1.6.0.
fix
Use: from nvidia_nat.config import AgentConfig
error ModuleNotFoundError: No module named 'nvidia_nat'
cause Library not installed or installed in wrong environment.
fix
Run: pip install nvidia-nat
error RuntimeError: NEMO_API_KEY environment variable not set
cause The library requires an API key for authentication.
fix
Set the environment variable, e.g., export NEMO_API_KEY='your_key' or pass api_key parameter to AgentConfig.
breaking In version 1.6.0, the AgentConfig class is no longer defined in nvidia_nat directly; it must be imported from nvidia_nat.config. Old imports (from nvidia_nat import AgentConfig) raise ImportError.
fix Update imports to 'from nvidia_nat.config import AgentConfig'.
deprecated The method 'agent.call()' is deprecated in 1.6.0 and will be removed in 2.0. Use 'agent.chat()' instead.
fix Replace agent.call() with agent.chat().
gotcha CUDA is required for GPU acceleration. If torch is installed without CUDA, the library silently falls back to CPU, which is extremely slow for large models.
fix Install PyTorch with CUDA: pip install torch --index-url https://download.pytorch.org/whl/cu118
pip install nvidia-nat[all]

Initialize a NemoAgent with an API key from environment variable and send a chat message.

import os
from nvidia_nat import NemoAgent
from nvidia_nat.config import AgentConfig

api_key = os.environ.get('NEMO_API_KEY', '')
config = AgentConfig(api_key=api_key)
agent = NemoAgent(config)
response = agent.chat("Hello, how are you?")
print(response)