Securing Agents with OpenShell and NemoClaw

Go from one model call to coordinated agents, grounded retrieval, deep planning, and safer execution. NemoClaw provides the reference stack; OpenShell provides the sandbox.

Start learning	Prepare the lab	Access models
Open the course	Launch NemoClaw on Brev	Open NVIDIA Build
Work through the four modules.	Keep it running for the live-agent exercises.	Create an API key when prompted.

About the course

An agent connects a model to tools, memory, and a routing decision that repeats until its task is done. This course builds that loop, then connects it to current frameworks. You’ll go from a single API call to agent coordination, grounded retrieval, deep planning, and safe deployment using OpenClaw, NVIDIA NemoClaw™, and OpenShell.

Every lesson is an editable browser exercise. Later modules use NemoClaw and OpenShell to run and inspect an agent.

Note

This repository releases the course and validation tools. NemoClaw and its runtime remain external dependencies.

Course path

Module 1: build the agent loop

Start with the observation-action loop. Add model calls, state, tools, and a clear stop condition.

Module 2: coordinate and ground the work

Add explicit routing, parallel work, retrieval over your own sources, and deeper planning.

Modules 3 and 4: run and constrain the agent

Connect the design to NemoClaw, keep an agent running, then use OpenShell to constrain tools, files, and network access.

What you will learn

• Build a basic agent loop and identify its core components.
• Implement reliable tool use and function calling within an agent system.
• Design and coordinate multi-agent systems using structured routing patterns.
• Utilize OpenShell to configure agent identities and ensure safe, sandboxed operations.
• Deploy and manage autonomous agents while building persistent skill libraries.

Take the course

1. Use the official NVIDIA DLI course page for NVIDIA-managed enrollment, then open the GitHub Pages course and work through Module 1.
2. Start the NemoClaw Brev launchable before Module 3, where the exercises drive a live agent.
3. Open NVIDIA Build and create an API key when a lesson asks for one.
4. Run the supplied examples, inspect their behavior, and change one input at a time.

The static browser site is available in English, Spanish, and Brazilian Portuguese. Its canonical entrypoint is the course source.

Run the course locally

Install Python 3.11+, Bash, Node.js 20+, pnpm through Corepack, and Chromium. Python 3.12 is the tested default.

git clone https://github.com/NVDLI/NemoClawDLI.git
cd NemoClawDLI
bash scripts/build/build_pages.sh public
python3 -m http.server -d public 8000

Open http://localhost:8000/nemoclaw/. Before installing Python packages, run python3 scripts/runtime/python_env_probe.py, use a virtual environment, and install the applicable pinned lock.

Contribute

Corrections, teaching ideas, runtime observations, and source leads are welcome through Issues. Broader questions belong in Discussions.

Start with CONTRIBUTING.md. Each patch needs an Issue, declared blast radius, validation evidence, and DCO signoff. See the Code of Conduct, support policy, and DCO.

Important

Report security issues through SECURITY.md. Do not place vulnerabilities, credentials, or access sessions in a public Issue or Discussion.

Verify a change

bash scripts/build/install-hooks.sh
python3 scripts/validation/release_gate.py --tier fast --no-write

Use --tier ship before release or after a cross-cutting contract change. The release test plan maps claims to evidence; browser setup lives in docs/lab_runtime_testing.md.

Code remains an untrusted proposal until checks and an authorized reviewer accept it. No contributor, maintainer, or agent may approve its own protected merge or release, and approval does not replace a missing control.

Agent guidance

AGENTS.md is the cross-harness contract. Directory-level SKILL.html beacons route people and agents to the relevant files and checks. The compliance suite supports review; it does not replace required security, license, review, or release controls.

Repository map

Path	Purpose
web/	Browser course, runtime, figures, materials, and dependency evidence
i18n/	Reviewed Spanish and Brazilian Portuguese overlays
scripts/	Build, validation, runtime, compliance, and material tooling
docs/	Design, test, deployment, security, and release contracts
SKILL.html	Repository-wide map

The Rapidly-Evolving Agentic Compliance Suite design note explains the repository’s workflow pattern and limits.

Security, dependencies, and release integrity

The threat model lives in docs/security-design.md. The browser dependency inventory, THIRD_PARTY_LICENSES.md, and THIRD-PARTY-NOTICES.md distinguish shipped code, tools, and sourced materials.

Governance and license

This NVIDIA-owned DLI course repository is approved for public release as a Full-OSS-Project, OSS Type I. That classification does not apply to the NemoClaw product.

RELEASE_STATUS.json records the public-safe release state. docs/release_playbook.md owns maintainer roles and publication; CHANGELOG.md records version history.

The project uses the Apache License 2.0. Contributions use the same inbound license with DCO signoff; no separate CLA is required.