Nemoclaw architecture

What is Nemoclaw?

NemoClaw is an open-source reference stack from NVIDIA for running AI agents more safely inside OpenShell sandboxes. It wraps supported agents — OpenClaw by default, alongside Hermes and LangChain Deep Agents Code — with guided onboarding, managed inference, network policy, and lifecycle operations driven from a CLI.

A reference stack is only useful if you can see how it is assembled. Which parts define the sandbox itself, and which configure the agent running inside it? How does one CLI support several different agents? Where do network policy and managed integrations actually get enforced — in the base image, the plugin layer, or the blueprint?

The map on this page is JigsawML's read of the NemoClaw repository. It separates the CLI, plugin, and blueprint layers from the sandbox image and the agents that run in it, and lets you open any area to see the code.

Key components in the Nemoclaw architecture

JigsawML identified and labelled these components automatically by analysing the Nemoclaw repository. Open the interactive map to expand any of them and read the underlying code.

• NemoClaw CLI

  The command-line entry point and the agent-specific aliases built on it.

• NemoClaw Plugin

  The plugin layer extending agent behaviour inside the sandbox.

• NemoClaw Blueprint

  The declarative stack definition describing a deployable configuration.

• Supported Agent

  Per-agent integration for OpenClaw, Hermes, and Deep Agents Code.

• Sandbox Base

  The base container image defining the sandbox an agent runs in.

• Automation PR

  Tooling for automated changes and pull-request workflows.

• Test Suites

  Tests covering the CLI and the stack's installation paths.

• Documentation Skills

  Documentation and the agent skills shipped with the stack.

• External Services

  OpenShell, inference endpoints, and other services the stack calls.

How this map was generated

JigsawML ingested the public repository, resolved its dependencies, and grouped the code into architectural components using AI agents and a knowledge graph. Nothing here was drawn by hand. The same process runs against private repositories and cloud accounts, and re-runs on every commit so the map never goes stale.

Other open-source architectures

• Axios architecturePromise-based HTTP client for the browser and node.js
• Flask architectureThe Python micro framework for building web applications
• Hono architectureAn ultrafast web framework built on Web Standards
• Instro architectureAn open-source Python library for scripting and automating test equipment
• Nanoclaw architectureA personal AI agent which runs securely in containers
• OpenClaw architectureA self-hosted gateway that connects chat apps to AI coding agents
• OpenHands architectureA self-hosted developer control center for coding agents and automations
• Pydantic architectureData validation library for Python
• Pyramid architectureThe Start Small, Finish Big, Stay Finished Framework
• Zod architectureA TypeScript-first validation library