An open-source AI agent platform published by NVIDIA. Running an always-on agent like OpenClaw directly on your own PC or workstation is unnerving — what if it deletes files, opens network connections on its own, or sends confidential data to a cloud model? NemoClaw answers those worries with three mechanisms: isolated execution (a sandbox), an allow-list policy layer, and routing that keeps sensitive data local. Its runtime component, NVIDIA OpenShell, is released under Apache 2.0 and runs on hardware you already have, from GeForce RTX to DGX Spark and DGX Station.
Key Features
- Isolated execution via the OpenShell sandbox: Agents run inside an isolated execution environment so they cannot damage host files or processes. Even a self-evolving agent that keeps adding its own skills stays contained inside the box. Policies can be updated while the agent is running, and a full audit trail is kept
- Fine-grained permissions from the policy engine: Constraints are applied across three layers — filesystem, network, and process — at the granularity of executable binary, destination, method, and path. The default is deny-everything, with developers approving only what is actually needed
- Out-of-process policy enforcement: Policy decisions happen outside the agent’s own process. Even if the agent is compromised or goes off the rails, it cannot rewrite the constraints imposed on it — this is the architectural cornerstone
- Privacy router: Sensitive data is processed locally with open models, and only reaches frontier cloud models such as Claude or GPT when policy permits. Which side handles a request is decided by cost and privacy policy, not by the agent’s preference
- Local inference with Nemotron: NVIDIA’s open Nemotron models run on your own GPU, so agents can operate without depending on external APIs. Options scale with your needs: RTX laptops, GeForce RTX, RTX PRO workstations, and DGX Spark / DGX Station
- Multiple agent implementations: Beyond the default OpenClaw, you can install Hermes (a self-improving agent built with Nous Research) or LangChain Deep Agents. The installer offers both an express setup with recommended defaults for DGX and WSL hosts and an interactive path for custom configuration
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Open source | $0 (Apache 2.0) | NemoClaw / OpenShell itself, sandbox, policy engine, privacy router, local inference with Nemotron |
| Runtime environment (separate) | Hardware cost | Purchasing GeForce RTX, RTX PRO, DGX Spark, DGX Station, and similar systems |
| Cloud model usage (separate) | Each vendor’s usage-based pricing | API charges for requests routed to frontier models such as Claude or GPT |
Pricing reflects information as of August 2026. Check the official site for the latest details.
Pros & Cons
✅ Pros
- The software itself is free and open source, and Apache 2.0 allows broad commercial use and modification
- You can declare what an agent must not do from outside the code, adding safety measures without rewriting the application
- You can keep confidential data local while still calling on high-performance cloud models when needed
- Audit logs let you review after the fact what the agent actually accessed
- You can choose among OpenClaw, Hermes, and LangChain Deep Agents as the agent implementation
⚠️ Cons
- It effectively assumes NVIDIA GPUs, and running local inference properly requires reasonably capable hardware
- You need to understand policy design and sandbox operations, which is heavy for someone who just wants to try an AI tool
- Support comes mainly through GitHub and the NVIDIA Developer Discord community, not a SaaS-style help desk
- Deployment targets Linux, DGX, and WSL; this is not a tool that lives entirely in a GUI
- It is still young, so documentation and the surrounding ecosystem are a work in progress
Comparison with Similar Services
| Criteria | NVIDIA NemoClaw | OpenClaw (standalone) | LangChain Deep Agents | E2B |
|---|---|---|---|---|
| Primary role | Safe execution platform for agents | The always-on agent itself | Agent-building framework | Cloud code execution sandbox |
| Isolated execution | OpenShell sandbox | Limited by default | Left to the implementer | Isolated on the cloud side |
| Policy control | Enforced out-of-process, deny by default | Configured case by case | Implemented case by case | Mostly runtime restrictions |
| Where it runs | Your own GPU / workstation | Local | Local and cloud | Cloud |
| Model handling | Local Nemotron plus routing to the cloud | Connected by the user | Any LLM | Execution only (models separate) |
| Price | Free (Apache 2.0) | Free | Free (OSS) | Usage-based |
Who Is It For
- People who want to use an always-on agent like OpenClaw but are uneasy about leaving file operations and outbound traffic unrestricted
- Teams handling internal documents or customer data who need a setup that keeps confidential information out of external APIs
- Developers with a GeForce RTX or DGX Spark on hand who want to run agents on local inference
- IT and security staff who need agent behavior to be explainable through audit logs
- Researchers who want to evaluate self-evolving, autonomous agents inside a box they can afford to break
Summary
NemoClaw is not a tool for building smarter agents; it is the foundation that lets the autonomous agents you already have keep running with confidence. The trio of sandboxing, deny-by-default policy, and the privacy router covers exactly the parts you cannot skip when putting an always-on agent to work in a business setting. It does assume NVIDIA GPUs and carries some infrastructure-side learning cost, but since the software is free under Apache 2.0, it is well worth following the GitHub instructions if you have a supported GPU at hand.