NVIDIA NemoClaw: The Open-Source AI Agent Platform, Explained

By Kortix Blog · Published September 27, 2026

When NVIDIA's NemoClaw entered the open-source AI agent platform conversation in March 2026, most coverage answered one question — that it exists — and left the more useful one open: what kind of tool it actually is and where it sits in the agent stack. NemoClaw is not a turnkey service you subscribe to; it is an open-source reference stack for running supported agents inside NVIDIA OpenShell sandboxes, and NVIDIA labels it an alpha project. Kortix (Suna) is an open-source AI management system that keeps agents, skills, company memory, and connectors in one git repo, runs each session in an isolated sandbox on its own branch, and lands work through a change request a human approves. This article explains NemoClaw on its own terms, then places the two on the same axes — license, where each runs, and how finished work actually lands — so a team evaluating open-source agent platforms can decide which fits.

What NVIDIA NemoClaw is

NVIDIA NemoClaw is a collection of open blueprints for building autonomous agents: domain-specialized, always-on AI systems that reason, plan, and act across real-world workflows. That definition comes from NVIDIA's own overview page, which frames NemoClaw as the bridge from prototype to "governed deployment" through runtime controls, model routing, skill execution, and observability. (nvidia.com/en-us/ai/nemoclaw)

Each NemoClaw blueprint bundles three NVIDIA components. NVIDIA Nemotron models (alongside other frontier models) provide the reasoning layer; NVIDIA NeMo handles specialization and optimization; and NVIDIA OpenShell enforces runtime policy controls over what the agent can access. The overview page describes these as "NVIDIA Agent Toolkit components." (nvidia.com/en-us/ai/nemoclaw)

The GitHub repository states the intent more precisely. NVIDIA NemoClaw is described there as "an open source reference stack for running supported AI agents more safely inside NVIDIA OpenShell sandboxes," providing guided onboarding, managed inference, network policy, managed integrations, snapshots, and lifecycle operations through the NemoClaw CLI. The supported agents are OpenClaw (the default), Hermes, and LangChain Deep Agents Code. (github.com/NVIDIA/NemoClaw)

Two qualifiers matter when evaluating NemoClaw. First, the repository is licensed under Apache 2.0, so it is OSI-approved open source. Second, NemoClaw is an alpha project — the README states that maintainers review issues, discussions, and pull requests "on a best effort basis without guaranteed response timelines." It is a reference stack for people who want to run and extend agents, not a managed product with a support SLA. (github.com/NVIDIA/NemoClaw)

How NemoClaw runs an agent

NemoClaw installs through a single shell script. The NVIDIA overview page lists the pattern curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash, with an optional NEMOCLAW_AGENT environment variable to select Hermes, LangChain Deep Agents Code, or OpenClaw (the default when the variable is omitted). (nvidia.com/en-us/ai/nemoclaw)

Once installed, the agent runs inside an OpenShell sandbox, and OpenShell is the security boundary. NVIDIA's overview page distinguishes the two: "NemoClaw is the full agent deployment package — models, harness, tools, and runtime," while "OpenShell is the secure runtime inside it that enforces what the agent can access: files, networks, credentials, and tools." The runtime adds managed inference, network policy, and lifecycle operations that are managed through the CLI. (nvidia.com/en-us/ai/nemoclaw, github.com/NVIDIA/NemoClaw)

NemoClaw is hardware-centric. NVIDIA positions the local, always-on compute it provides for agents on NVIDIA RTX Spark laptops, GeForce RTX PCs and laptops, RTX PRO workstations, and DGX Station or DGX Spark. The GitHub repository's prerequisites specify a supported DGX or Windows Subsystem for Linux (WSL) host. A team without that hardware cannot run NemoClaw as designed. (nvidia.com/en-us/ai/nemoclaw, github.com/NVIDIA/NemoClaw)

NemoClaw vs Kortix: runtime vs management system

The core distinction is the layer each product operates at. NemoClaw is a runtime and reference stack for executing individual agents under policy; Kortix is a management system for governing an entire agent workforce across a company. Both isolate agents in sandboxes, but they own different parts of the workflow.

Dimension	NVIDIA NemoClaw	Kortix (Suna)
What it is	Open-source reference stack / blueprints for running agents in OpenShell sandboxes	AI management system — agents, skills, memory, and connectors in one git repo
License	Apache 2.0 (OSI open source)	Elastic License 2.0 (source-available, not OSI)
Where it runs	NVIDIA hardware: RTX Spark, GeForce RTX, RTX PRO, DGX Station/Spark; supported DGX or WSL host	Any model, any infrastructure — cloud, VPC, or on-prem; self-host free
How work lands	Agent executes inside an OpenShell sandbox under runtime policy, network, and observability controls	Agent commits on its own branch, then opens a change request a human reviews and merges
Configuration	Blueprints plus the NemoClaw CLI (guided onboarding, lifecycle operations)	Files in a git repo you own (kortix.yaml), versioned and diffable
Model choice	Nemotron and other models, with model routing under policy	Any provider with your own API keys
Maturity	Alpha — best-effort maintainer review	Self-host free; managed cloud $40/seat/month plus usage (vendor-listed)

The "how work lands" row is the sharpest contrast. In NemoClaw, the guardrail is the sandbox itself: OpenShell enforces which files, networks, and credentials the agent can touch while it runs. In Kortix, the guardrail is a review step: each session runs an agent on its own branch, and what the agent produces reaches the main branch only through a change request that a human approves. One governs execution; the other governs the finished work. (github.com/NVIDIA/NemoClaw, kortix.com/docs, github.com/kortix-ai/suna)

Is NemoClaw actually open source?

Yes. NemoClaw is licensed under Apache 2.0, which is an OSI-approved open-source license, and the GitHub repository confirms it ("Apache 2.0. See LICENSE"). Apache 2.0 permits commercial use, modification, and redistribution with relatively permissive terms. (github.com/NVIDIA/NemoClaw)

Kortix is different, and the distinction is worth stating precisely. Kortix is licensed under Elastic License 2.0, which is source-available rather than OSI open source. The license permits use, modification, and redistribution, but forbids providing the software to third parties as a hosted or managed service that exposes "any substantial set of the features or functionality of the software." In practice you can read the code, fork it, and self-host it, but you cannot stand up a competing managed Kortix service. (github.com/kortix-ai/suna/blob/main/LICENSE)

The practical distinction for a buyer: NemoClaw's Apache-2.0 terms let you take the stack and build a hosted offering around it; Kortix's Elastic 2.0 terms reserve the hosted/managed-service right to Kortix. If the question is "which one can I legally resell as a service," the answer is NemoClaw, not Kortix. Both, however, let you self-host and run your own data and models.

The "claws" context

NemoClaw's name leans on a broader movement around "claws" — open-source AI tools that run locally on a user's machine and perform multi-step tasks. The most visible claw is OpenClaw, which was first named Clawdbot, then Moltbot, and which "captivated Silicon Valley" for running autonomously on personal computers before OpenAI acquired the project and hired its creator. (wired.com)

NVIDIA's positioning of NemoClaw is tied to that momentum. CNBC's report on the launch notes that Jensen Huang called OpenClaw "the most important software release probably ever," and frames NemoClaw as NVIDIA tapping the claws trend for enterprise customers. NemoClaw's default supported agent is OpenClaw, and NVIDIA states on its overview page that it "continues contributing to the OpenClaw project." (cnbc.com, nvidia.com/en-us/ai/nemoclaw)

The claws trend also carries enterprise caution that shapes why NemoClaw exists. WIRED reported in March 2026 that some tech companies, including Meta, have asked employees to refrain from using OpenClaw on work computers because of the agents' unpredictability and security risk, and that a Meta employee who oversees AI safety and alignment publicly described an agent going rogue on her machine and mass-deleting emails. NVIDIA's pitch — an open-source stack with added security and privacy controls and OpenShell sandboxing — is a direct response to that caution. (wired.com)

WIRED's original report also describes the launch mechanics: ahead of GTC in March 2026, NVIDIA pitched NemoClaw to Salesforce, Cisco, Google, Adobe, and CrowdStrike, and because the platform is open source, partners would likely get free early access in exchange for contributing. Companies would be able to dispatch agents regardless of whether their products run on NVIDIA chips, and the platform was expected to include security and privacy tools. WIRED noted it was unclear whether those conversations produced official partnerships. (wired.com)

When to choose which

Choose NemoClaw when the constraint is execution safety on NVIDIA hardware. If your team already runs NVIDIA RTX, RTX PRO, or DGX systems and wants an Apache-2.0 reference stack to run an individual agent — OpenClaw, Hermes, or LangChain Deep Agents Code — under OpenShell policy controls, NemoClaw fits. It is the right call when the deliverable is a governed, always-on agent on local NVIDIA compute. The trade-offs: it is alpha software with best-effort support, and it is tied to NVIDIA hardware and WSL/DGX hosts. (github.com/NVIDIA/NemoClaw, nvidia.com/en-us/ai/nemoclaw)

Choose Kortix when the constraint is managing a whole workforce across any infrastructure. If you want many specialist agents sharing one company memory, a git-native configuration you can diff and audit, and human-approved change requests as the gate on what reaches production, Kortix is aimed at that. Kortix is model- and infrastructure-agnostic — any provider with your own API keys, on cloud, VPC, or on-prem — with self-hosting free and a managed cloud listed at $40/seat/month plus usage. The trade-offs: it is source-available under Elastic 2.0 rather than OSI open source, and the hosted-service right is reserved to Kortix. (github.com/kortix-ai/suna, kortix.com/docs)

The two are not strictly competitors so much as different layers. A team could run an agent harness on NemoClaw and still want a management layer that turns a fleet of agents into reviewed, versioned work — which is where a git-native system like Kortix operates. For a fuller comparison of open-source agent platforms for teams, see our open-source AI agent platform guide; for the closed-source alternatives Kortix positions against, see open-source Claude Cowork alternatives.

FAQ

Is NemoClaw free?

NemoClaw is free in the software-license sense: it is Apache 2.0 open source, so you can use, modify, and redistribute it without a license fee. It is not free in the infrastructure sense — you must supply the NVIDIA hardware it runs on (RTX Spark, GeForce RTX, RTX PRO, or DGX systems, with a supported DGX or WSL host). It is also alpha software with no guaranteed support timelines. (github.com/NVIDIA/NemoClaw)

Does NemoClaw require NVIDIA hardware?

Yes. The GitHub repository lists a supported DGX or Windows Subsystem for Linux (WSL) host as a prerequisite, and NVIDIA's overview page describes the local compute as running on NVIDIA RTX Spark laptops, GeForce RTX PCs and laptops, RTX PRO workstations, and DGX Station or DGX Spark. There is no stated non-NVIDIA hardware path. (github.com/NVIDIA/NemoClaw, nvidia.com/en-us/ai/nemoclaw)

Is NemoClaw the same as OpenClaw?

No. OpenClaw is a specific autonomous agent (formerly Clawdbot, then Moltbot) that OpenAI acquired. NemoClaw is NVIDIA's open-source reference stack for running supported agents — OpenClaw is its default, alongside Hermes and LangChain Deep Agents Code — inside OpenShell sandboxes. NVIDIA also continues contributing to OpenClaw, but NemoClaw and OpenClaw are separate things. (nvidia.com/en-us/ai/nemoclaw, wired.com)

Is Kortix open source?

Kortix is source-available, not OSI open source. Its license is Elastic License 2.0, which lets you read, fork, and self-host the code but forbids offering it to third parties as a hosted or managed service. NemoClaw, by contrast, is Apache 2.0, which is OSI-approved. If "open source" means an OSI license, NemoClaw qualifies and Kortix does not. (github.com/kortix-ai/suna/blob/main/LICENSE, github.com/NVIDIA/NemoClaw)

What is the difference between a runtime and a management system?

A runtime executes an agent and enforces guardrails while it runs — NemoClaw's OpenShell controls what files, networks, and credentials an agent can access. A management system governs the whole agent workforce across its lifecycle — Kortix keeps agents, skills, memory, and connectors in one git repo and lands work only through a change request a human approves. One is about safe execution; the other is about governed, reviewable work. (github.com/NVIDIA/NemoClaw, kortix.com/docs)