Nvidia’s NemoClaw Brings Security and Governance to the Open-Source Agent Stack

Quick Facts

• Nvidia announced NemoClaw at GTC 2026, adding security, privacy, and governance layers to the open-source OpenClaw agent platform via a single install command.
• The companion model Nemotron 3 Ultra is a 550-billion-parameter model that Nvidia says delivers up to 5x faster inference and 30% lower cost than comparable open frontier models.
• Seventeen enterprise software platforms, including Salesforce, SAP, CrowdStrike, and Microsoft, have joined the NemoClaw partner ecosystem.

Nvidia launched NemoClaw at GTC 2026, its answer to the security problems that have prevented enterprises from deploying autonomous AI agents at scale. The product installs Nvidia’s Nemotron models and its new OpenShell runtime onto OpenClaw, the fastest-growing open-source project in GitHub history, with a single command.

OpenClaw’s rise was fast and chaotic. Originally launched in November 2025 as a side project called ClawdBot, it accumulated more than 321,000 GitHub stars in 60 days, surpassing React’s record. But an unsecured database that allowed anyone to impersonate any agent on the platform prompted Meta and other large technology companies to ban it from corporate machines entirely.

NemoClaw is Nvidia’s attempt to fix that. Jensen Huang told the GTC 2026 audience that OpenClaw and Claude Code had sparked an agent inflection point, and that every company now needs an OpenClaw strategy. Huang compared claws to Linux, Kubernetes, and HTML as foundational tools.

How the Stack Works

NemoClaw is built on three layers. The first is OpenShell, an open-source security runtime that enforces behavior from outside the agent process using kernel-level isolation. It adds deny-by-default filesystem policies via Landlock LSM, network namespace rules for egress control, and seccomp filter profiles that restrict system calls the agent can invoke. Kari Briski, Nvidia’s VP of Generative AI Software, described OpenShell as the missing infrastructure layer beneath claws.

The second layer is a Privacy Router that sends sensitive data to local models first and only routes information to frontier models like Claude when policy permits. The routing decision is made by the harness, not by the agent itself. The third layer is native support for Nemotron models, which run on-premises so data never leaves a company’s own infrastructure.

Early benchmarks show NemoClaw agents achieving a 2x improvement in planning speed compared to standard LangChain implementations, which Nvidia attributes to a mechanism called Memory-Mapped Context. This allows agents to retain goal-state across multiple hardware cycles without redundant reprocessing.

Partner Ecosystem

Seventeen enterprise software platforms have committed to the Nvidia Agent Toolkit, which underpins NemoClaw. Salesforce is integrating Nemotron models into Agentforce for service, sales, and marketing use cases. SAP is using the toolkit with NeMo to power AI agents through Joule Studio on its Business Technology Platform.

CrowdStrike unveiled a Secure-by-Design AI Blueprint at GTC that embeds its Falcon platform directly into OpenShell, adding AI policy enforcement, endpoint protection, cloud runtime monitoring, and identity-based governance. Microsoft is working with Nvidia to deliver native Windows AI agents built on OpenShell’s security primitives.

Nvidia engineer Nader Khalil framed the enterprise agent problem as one of orchestration, not model development, stating that an agent is an LLM and a harness. NemoClaw is Nvidia’s bet that the harness is where enterprise contracts will be won.

For founders and executives evaluating agent infrastructure, NemoClaw shifts the calculus. Kernel-level sandboxing and on-premises model routing address the two objections that have kept legal and security teams from approving autonomous agent deployments: data leaving the environment and agents acting outside defined boundaries. With 17 major software vendors now building on the same stack, the ecosystem pressure to standardize on NemoClaw will be difficult to ignore.

Read more: Why Nvidia’s NemoClaw signals the true enterprise agent era