NVIDIA NeMo Agent Toolkit
Open-source toolkit for connecting and optimizing teams of AI agents — profiling, evals, and deployment included.
About
The NeMo Agent Toolkit (NAT) is NVIDIA's open-source answer to the 'day two' problems of agent development: turning a demo agent into something you can profile, evaluate, observe, and serve. You define agentic workflows in YAML configs and run them with the aiq CLI — simple chat apps up to hierarchical setups where one agent recursively delegates to others as tools. It plugs into LangChain and LlamaIndex, exposes workflows as MCP servers, and ships blueprints like AI-Q (a hybrid research agent) and an alert-triage agent.
Release 1.1.0 added full MCP support, a ReWOO agent that trims tool output from context to cut token costs, and deep integration with Weights & Biases Weave for tracing. The positioning is less 'another agent framework' and more the glue between the frameworks you already use and production. The catch is youth: it hit 1.0 in 2026, the ecosystem examples lean NVIDIA (NIM, Nemotron), and at ~2,600 stars the community is still forming.
What it can do
- YAML-defined agentic workflows run via the `aiq` CLI
- Hierarchical agents that delegate to other agents as tools
- MCP server support — expose any workflow as an MCP server
- Built-in profiling, evaluation, and observability hooks (Weave, Phoenix)
- Blueprints: AI-Q research agent, alert triage, ReWOO, and more
- Serve workflows as REST APIs with a bundled web UI
Setup
- Install Python 3.11+ and the toolkit:
pip install nemo-agent-toolkit(or clone the repo anduv pip install -e .). - Export your model key (e.g.
NVIDIA_API_KEYfor NIM, or OpenAI). - Pick an example config from
examples/or write your own workflow YAML. - Run it:
aiq run --config_file <config>.yml --input "your task". - Serve it as an API:
aiq serve --config_file <config>.yml, optionally launching the bundled UI.
Prerequisites
Python 3.11+; a model provider key (NVIDIA NIM, OpenAI, or others); Docker optional for containerized runs.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: GitHub API shows 2,661 stars, Apache-2.0, pushed Oct 15 2026; release notes document 1.0/1.1.0 with MCP support and Weave integration; SiliconANGLE's March 2026 launch coverage corroborates the positioning and partner list. Strengths: first-party NVIDIA tooling, genuinely open, tackles real productionization gaps (profiling, evals, serving). Docked for: (1) young project — 1.0 landed in 2026, long-term direction unproven; (2) examples and defaults tilt toward NVIDIA's own stack; (3) smaller community than LangGraph/CrewAI. License: Apache-2.0.
★ 2,661 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 10.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 7: starScore from 2,661 stars on NVIDIA/NeMo-Agent-Toolkit
- · Maintenance 10: pushed 2026-10-15 (recency 10); 1.1.0 release with MCP support and Weave integration
- · Openness 10: Apache-2.0 + official docs verified (docs.nvidia.com)
- · Manual 8: release notes and SiliconANGLE launch coverage corroborate the productionization positioning; docked for youth (1.0 in 2026), NVIDIA-stack tilt in examples, smaller community than LangGraph/CrewAI
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.