NVIDIA NeMo Agent Toolkit
Python AI agent toolkit for self-hosted MCP and A2A servers, workflow profiling and evaluation. Apache 2.0 licensed; its NIM example uses hosted models.
NVIDIA NeMo Agent Toolkit is an open-source Python library for developers building self-hosted servers that expose AI tools and workflows through MCP or A2A. It also measures and evaluates agent workflows across frameworks. Its introductory NVIDIA NIM workflow uses hosted inference and requires an NVIDIA account and API key.
Teams can keep LangChain, LlamaIndex, CrewAI, Microsoft Semantic Kernel or Google ADK, as well as custom Python agents. The toolkit adds reusable components and instrumentation around those frameworks rather than requiring a replacement stack.
MCP support lets agents call remote tools and publish their own tools or workflows as servers, including through FastMCP. A2A support lets workflows delegate work to remote agents or expose themselves as discoverable agents. A chat interface helps developers inspect responses and debug workflows.
The profiler tracks token usage and timings at workflow, agent and tool level. Built-in evaluation checks accuracy, while prompt and parameter optimization compares configurations. Execution tracing integrates with LangSmith, Phoenix, Weave and Langfuse, with OpenTelemetry compatibility. Graph-based workflows can use parallel execution, speculative branches and priority routing.
The license is Apache 2.0. Optional CLI telemetry sends command outcomes and timing data to NVIDIA; it excludes configuration contents, command output and identifying information.
