An open-source orchestration and optimization toolkit for AI Agents that connects, observes, and tunes multi-agent systems in a framework-agnostic manner, with MCP/A2A protocol support, end-to-end profiling, offline evaluation, and RL fine-tuning.
NVIDIA NeMo Agent Toolkit (PyPI package nvidia-nat) is a meta-orchestration layer for Agentic systems released by NVIDIA, designed with framework-agnosticism as its core principle. Rather than replacing frameworks like LangChain, CrewAI, or AutoGen, it provides unified workflow definition, protocol bridging, and performance optimization on top of them.
The workflow engine is driven by YAML declarative configuration, defining Agent behavior through a three-part structure of functions, llms, and workflow. It includes built-in workflow types like ReAct Agent and supports component reuse across Agents, tools, and workflows.
On the protocol side, NAT supports both MCP (acting as Client to consume remote tools or as Server/FastMCP Server to publish tools) and A2A (acting as Client to delegate tasks or as Server to publish discoverable Agents with authentication), enabling bidirectional interconnection of tools and Agents.
For observability, it provides end-to-end profiling from workflow level down to token level, with native LangSmith integration and support for Phoenix, Weave, Langfuse, and OpenTelemetry. The Dynamo Runtime Intelligence module automatically infers latency sensitivity for cache control, load-aware routing, and priority-aware scheduling.
The evaluation and optimization system covers offline evaluation, automatic hyperparameter and prompt optimization, reinforcement learning fine-tuning for specific Agents, and framework-agnostic Agent Performance Primitives (APP) supporting parallel execution, speculative branching, and node-level priority routing.
Built as a Monorepo with independent sub-packages under packages/, it uses uv for dependency management and extras for on-demand installation. Supports Docker containerized deployment with nat run as the CLI entry point. Security features include built-in PII protection and MCP/A2A authentication.
Installation:
pip install nvidia-nat # Core package
pip install "nvidia-nat[langchain]" # With framework plugin
pip install "nvidia-nat[most]" # All optional dependencies
Requires Python ≥ 3.11, < 3.14. Latest release v1.6.0, development branch develop (1,262+ commits), Apache-2.0 license.