NAT

NVIDIA NeMo Agent Toolkit (NAT) — install, create workflows, add tools, run agents, evaluate performance, and publish as A2A/MCP servers. Use when: (1) insta...

Security Scan

The skill's declared purpose (NVIDIA NeMo Agent Toolkit) aligns with its instructions and required binaries; it is internally coherent but can require additional optional credentials and can publish network-accessible agents, so review configs and secrets before use.

Purpose & Capability

Name/description, the required binaries (nat, pip, uv), and the primary credential (NVIDIA_API_KEY) are consistent with installing and using the NVIDIA NeMo Agent Toolkit. Examples referencing S3, OpenAI, Bedrock, etc., are plausible optional integrations for this toolkit rather than unexplained or unrelated demands.

Instruction Scope

SKILL.md stays on-topic: install, configure, run workflows, create custom tools, and publish A2A/MCP servers. It does instruct the user to start network services (nat a2a serve) and to place credentials into workflow YAMLs (e.g., ${S3_ACCESS_KEY}), so workflows can cause the agent to access external networks and third-party services. This is expected for an agent toolkit but is a material operational/privacy consideration.

Install Mechanism

Instruction-only skill (no install spec) that recommends standard pip installs (via 'uv pip install' or 'pip install'). No unusual download URLs or archive extraction are present in the provided docs.

Credentials

The skill declares NVIDIA_API_KEY as primaryEnv which matches NIM usage. However the documentation shows many optional integrations (S3 access_key/secret, OpenAI/Azure/Bedrock keys, NAT_PUBLIC_BASE_URL, etc.). Those are not required by default but will be needed if you enable corresponding components—review any workflow configs for embedded secrets and be cautious about using broad credentials.

Persistence & Privilege

Skill is not marked always:true and is user-invocable; it does not request persistent system-wide privileges or modify other skills' configurations in the docs. Running A2A servers will open network endpoints by design, which is a normal capability for this toolkit.

Assessment

This skill appears to be what it claims: documentation for NVIDIA's NeMo Agent Toolkit. Before installing or running workflows, review any workflow YAMLs and example configs for embedded secrets (S3 keys, OpenAI/Azure keys, NAT_PUBLIC_BASE_URL) and avoid putting long-lived credentials into public configs. If you publish an A2A agent (nat a2a serve), understand you are opening an HTTP endpoint—use authentication and limit network exposure. When installing packages via pip, prefer official releases (GitHub/NVIDIA docs) and inspect any third-party example code you copy (custom tools can run arbitrary Python and use boto3/requests). If you need a higher-assurance review, provide the specific workflow YAMLs or example code you plan to run so those can be checked for secret leakage or unexpected network destinations.

NVIDIA NeMo Agent Toolkit (NAT)

A flexible library for connecting enterprise agents to data sources and tools across any framework.

• Repo: https://github.com/NVIDIA/NeMo-Agent-Toolkit
• Docs: https://docs.nvidia.com/nemo/agent-toolkit/latest/

Installation

# Core (pick one)
uv pip install nvidia-nat        # recommended
pip install nvidia-nat

# With framework extras
uv pip install "nvidia-nat[langchain]"    # LangChain/LangGraph
uv pip install "nvidia-nat[llama-index]"  # LlamaIndex
uv pip install "nvidia-nat[crewai]"       # CrewAI
uv pip install "nvidia-nat[mcp]"          # MCP
uv pip install "nvidia-nat[a2a]"          # A2A
uv pip install "nvidia-nat[mem0ai]"       # Mem0 memory
uv pip install "nvidia-nat[eval,profiling]"  # Eval + profiling

# Verify
nat --help && nat --version

For development install from source, see references/install-from-source.md.

Quick Start

export NVIDIA_API_KEY=<key_from_build.nvidia.com>

Create workflow.yml:

functions:
  wikipedia_search:
    _type: wiki_search
    max_results: 2

llms:
  nim_llm:
    _type: nim
    model_name: meta/llama-3.1-70b-instruct
    temperature: 0.0

workflow:
  _type: react_agent
  tool_names: [wikipedia_search]
  llm_name: nim_llm
  verbose: true
  parse_agent_response_max_retries: 3

nat run --config_file workflow.yml --input "List five subspecies of Aardvarks"

Workflow Configuration Structure

Four main YAML sections:

Section	Purpose
functions	Tools (web search, calculators, custom)
llms	LLM provider configs (NIM, OpenAI, Azure, Bedrock)
embedders	Embedding models for vector storage
workflow	Agent type + wiring of tools and LLMs

Agent Types (_type in workflow)

• react_agent — Reasoning and acting
• reasoning_agent — Advanced reasoning
• rewwo_agent — Reasoning Without Observation
• responses_api_agent — OpenAI Responses API
• tool_calling_agent — Direct tool calling
• automatic_memory_wrapper_agent — Adds memory
• router_agent — Routes to different workflows
• sequential_executor — Sequential tool execution

Built-in Tools (_type in functions)

wiki_search, webpage_query, tavily_internet_search, arxiv_search, current_datetime, calculator, text_file_ingest, and many more framework-specific tools.

List all available components:

nat info components -t function      # Tools
nat info components -t llm_provider  # LLMs
nat info components -t embedder      # Embedders

Common CLI Commands

# Run workflow
nat run --config_file workflow.yml --input "question"

# Override params without editing YAML
nat run --config_file workflow.yml --input "question" \
  --override llms.nim_llm.temperature 0.7 \
  --override llms.nim_llm.model_name meta/llama-3.3-70b-instruct

# Create new workflow template
nat workflow create --workflow-dir examples my_workflow

# Evaluate
nat eval --config_file eval_config.yml

# Profile
nat profiler --config_file workflow.yml --input "test"

# Red team
nat red-team --config_file workflow.yml

# Workflow management
nat workflow reinstall my_workflow
nat workflow delete my_workflow

Custom Tools and Function Groups

For creating custom tools, function groups, and advanced patterns, see:

• references/custom-tools.md — Writing custom functions, registration, and installation
• references/function-groups.md — Shared config, namespacing, include/exclude, access levels

A2A Server

Publish workflows as A2A agents for discovery and invocation by other A2A clients.

# Start A2A server
nat a2a serve --config_file workflow.yml

# Discover agent
nat a2a client discover --url http://localhost:10000

# Call agent
nat a2a client call --url http://localhost:10000 --message "What is 42 * 67?"

For full A2A configuration (auth, concurrency, Kubernetes), see references/a2a-server.md.

Examples

The repo includes examples organized by category: Getting Started, Agents, Advanced Agents, Control Flow, Frameworks, MCP/A2A, Evaluation, and more. See references/examples.md for the full catalog and how to run them.

# Run any example
uv pip install -e examples/<example_directory>
nat run --config_file examples/<example_directory>/configs/config.yml --input "test"