NVIDIA NemoClaw

Scroll sideways for more →

NVIDIA NemoClaw is a collection of open blueprints (reference architectures) for building custom autonomous AI agents that packages three moving parts, a model, an agent harness, and a secure runtime, into a single starting point for teams building specialized, domain-specific agents.^[1][2] NVIDIA announced NemoClaw on March 16, 2026 at GTC 2026, alongside the NVIDIA OpenShell runtime it depends on, as a way to install Nemotron models and OpenShell "in a single command" and add privacy and security controls to the OpenClaw agent platform.^[19] A LangChain co-developed "NemoClaw for LangChain Deep Agents" blueprint followed on July 8, 2026, and NemoClaw was a headline item at NVIDIA's SIGGRAPH 2026 keynote on July 20, 2026, where the company showed NemoClaw agents extending into physics simulation and 3D content tools.^[3][4][6] Rather than a single product, NemoClaw is a set of opinionated, ready-to-run stacks that pair an open Nemotron model with an agent framework and NVIDIA's OpenShell runtime, installed with hardened security defaults so an enterprise can stand up a working agent and then tune it for its own workloads.^[1][10]

What is NemoClaw

NVIDIA describes NemoClaw as "open blueprints for building custom autonomous agents, packaging the model, harness and runtime together as a starting point for teams building specialized, domain-specific agents."^[2] Its product page frames the same idea for practitioners: "a collection of open blueprints for building autonomous agents: domain-specialized, always-on AI systems that reason, plan, and act across real-world workflows," running locally on NVIDIA RTX, DGX Spark, and DGX Station hardware.^[1] The repository is blunter: NemoClaw is "an open source reference stack for running always-on AI agents more safely inside NVIDIA OpenShell sandboxes," providing "guided onboarding, a hardened blueprint, routed inference, network policy, and lifecycle management through a single CLI."^[20]

The key word is blueprint. NemoClaw is a reference implementation, not a framework you assemble from scratch. Each blueprint answers the three questions that usually stall an AI agent project: which model runs the reasoning, which harness (the loop that plans, calls tools, and manages memory) drives the agent, and which runtime safely executes what the agent decides to do. NVIDIA supplies a default answer for each layer, wired together and installable in effectively one command, then leaves every layer swappable.^[1][10] The blueprints are aimed at "always-on" or long-running agents: systems that persist in the background, check a task list on an interval, and either act or wait for the next cycle, rather than one-shot chatbots that respond only when prompted.^[10]

NemoClaw is released under the Apache 2.0 license, is written mostly in TypeScript, and had accumulated roughly 22,000 GitHub stars and 3,000 forks by July 31, 2026, with the tag v0.0.98 current.^[20] NVIDIA labels the project alpha and "early preview" and states plainly that maintainers review issues, discussions, and pull requests on a best-effort basis with no guaranteed response times.^[20][21]

NemoClaw is not the sandbox itself. NVIDIA's own ecosystem documentation splits the three projects by scope: OpenClaw is "the assistant: runtime, tools, memory, and behavior inside the container," OpenShell is "the execution environment: sandbox lifecycle, network, filesystem, and process policy, inference routing," and NemoClaw is "the NVIDIA reference stack on the host," the CLI, plugin, versioned blueprint, and messaging configuration that make the other two work together in a repeatable way.^[22] Running openshell sandbox create --from openclaw already produces a working sandboxed OpenClaw; what NemoClaw adds on top is hardening, onboarding automation, and lifecycle tooling.^[22]

Background: NVIDIA's agentic-AI push and the LangChain partnership

NemoClaw sits inside NVIDIA's broader 2026 push to make agentic AI run locally on its own hardware. The building blocks predate it: the open Nemotron model family for reasoning, the NVIDIA NeMo Agent Toolkit for connecting models to tools, skills, and data, and the DGX Station and DGX Spark desktop systems that put data-center-class inference on a workstation. NemoClaw is the layer that packages those pieces into a deployable agent instead of a parts bin.^[1][9]

The public start was GTC 2026 in San Jose on March 16, 2026, where NVIDIA announced NemoClaw and OpenShell together and pitched the pair as the way to make "self-evolving, autonomous AI agents, or claws, more trustworthy, scalable and accessible."^[19] Jensen Huang framed it around the community project rather than NVIDIA's own stack: "OpenClaw is the operating system for personal AI. This is the moment the industry has been waiting for." OpenClaw creator Peter Steinberger, quoted in the same release, described the collaboration as "building the claws and guardrails that let anyone create powerful, secure AI assistants."^[19] The NemoClaw repository was created a day earlier, on March 15, 2026, and NVIDIA's support matrix records the project's start date as 2026-03-16.^[20][21]

The LangChain blueprint came later. On July 8, 2026, the two companies announced the "NemoClaw for LangChain Deep Agents" blueprint, described as a reference architecture "for building open agent systems with benchmark-leading performance and more than 10x lower inference costs."^[3][4][5] LangChain co-founder and CEO Harrison Chase framed the collaboration around tuning the whole system rather than swapping in a bigger model: "The way to build better agents is to keep improving the system around the model. Memory, tool use, evaluation, and model behavior compound when teams can tune them together."^[4] Huang put it in product terms: "Super agents have arrived. With an open model like NVIDIA Nemotron, a LangChain harness, the NVIDIA OpenShell runtime, and a company's own data, every enterprise can build custom agents."^[4]

Twelve days later, at SIGGRAPH 2026 (July 20, keynote at 3:45 p.m. PT), NVIDIA folded NemoClaw into its graphics-and-simulation story, showing that the same agents could be pushed beyond text and code into 3D simulation and asset workflows through new Omniverse libraries.^[2][6]

Components

Every NemoClaw blueprint is built from the same four-layer stack. Only the harness (and, with it, the tuned model size) changes between blueprints.

Scroll sideways for more →

Model: Nemotron 3

The reasoning engine is NVIDIA's open Nemotron 3 family. The flagship Deep Agents blueprint runs on Nemotron 3 Ultra, which NVIDIA describes as "a frontier 550-billion-parameter open model" optimized to run on DGX Station GB300 systems; NVIDIA's technical materials characterize it as a hybrid Mamba-Transformer mixture-of-experts design with roughly 55 billion active parameters, pretrained in the NVFP4 format for efficient long-running reasoning.^[2][11][12] At SIGGRAPH, NVIDIA said Nous Research likewise fine-tuned Nemotron 3 Ultra for its Hermes Agent harness and adopted it for production workloads.^[2] Because the model layer is swappable, blueprints can also run smaller Nemotron variants: NVIDIA's Hermes walkthrough demonstrates the 120B Nemotron 3 Super, and any blueprint can point at a self-hosted model served through NVIDIA NIM or vLLM.^[8] Nemotron 3 has its own detailed article; NemoClaw simply treats it as the default, tuned model.

The Super tier is the one most NemoClaw users will actually meet, because it is the default in NVIDIA's own Hermes walkthrough. Its model card gives 120 billion total parameters with 12 billion active, a "LatentMoE" architecture combining Mamba-2, mixture-of-experts and attention blocks with multi-token prediction, a context window of up to 1 million tokens, and a minimum requirement of eight H100-80GB GPUs to self-host.^[24] The "a12b" in the model identifier nvidia/nemotron-3-super-120b-a12b is the active-parameter count, not a quantization label: NVFP4 and BF16 builds are published as separate repositories.^[24] Nemotron 3 models are not Apache-licensed like NemoClaw itself; they ship under the NVIDIA Nemotron Open Model License.^[24]

Harnesses: OpenClaw, Hermes, and Deep Agents

A harness is the control loop around the model: it plans, calls tools, manages memory and state, and decides the next step. NemoClaw ships three, and none of the three was written by NVIDIA, which supplies the packaging, the runtime, and the model instead.^[4][7]

• OpenClaw (default). OpenClaw is the community "always-on" agent that gives the whole family its name. NVIDIA describes it as an independent, MIT-licensed open-source project, created by developer Peter Steinberger and separate from NVIDIA, for self-hosted, persistent assistants that "check their task list, evaluate what needs action, and either act or wait for the next cycle."^[10] At the implementation level NemoClaw is packaged as an OpenClaw plugin plus an onboard setup wizard.^[13] It is onboarded with plain nemoclaw onboard, no agent flag required.^[21]
• Hermes Agent (Nous Research). The Hermes harness contributes a self-evolving skills-and-memory loop: when a user teaches it a preferred output format, "Hermes writes a SKILL.md to the filesystem" with structured scaffolding, and those skills persist across deployments via snapshot and restore, so "the more users work with the agent, the better it gets."^[8] Hermes is MIT-licensed, written in Python, and maintained by Nous Research at github.com/NousResearch/hermes-agent.^[27] NemoClaw treats it as a first-class agent with its own CLI (nemohermes), Dockerfile, manifest and end-to-end test jobs, onboarded either as nemohermes onboard or nemoclaw onboard --agent hermes; NVIDIA nonetheless notes that "production parity with OpenClaw is not yet asserted."^[21]
• LangChain Deep Agents. LangChain's Deep Agents framework adds explicit planning, parallel subagents with isolated context, and a virtual filesystem for persisting knowledge across sessions; it is the harness behind the headline enterprise blueprint and cost claim.^[3][18] In NemoClaw it appears as "LangChain Deep Agents Code," a terminal-oriented coding harness with the binary dcode, run as a managed harness whose sandbox, MCP and shell overrides NemoClaw rejects.^[21]

Runtime: OpenShell

NVIDIA OpenShell is the secure-by-design runtime shared by all three harnesses, and it is what makes an autonomous agent safe enough to leave running. NVIDIA positions it as a governance layer that sits "between your agent and your infrastructure," providing programmable system and network isolation, a policy engine that enforces constraints across the filesystem, network, and process layers (declared in a version-controlled policy file), credential brokering so the agent never handles raw tokens, live policy updates, and full audit trails.^[1][9] Its Privacy Router "keeps sensitive context on-device with local open models and routes to frontier models like Claude and GPT only when policy allows," which is the mechanism behind NemoClaw's model-routing pitch of getting "the best of both" local and cloud models under defined controls.^[1][9] OpenShell is documented as part of the NVIDIA Agent Toolkit and is also surfaced on the DGX Station for Windows materials.^[9][14]

The component count has moved since launch. NVIDIA's March 2026 OpenShell blog described three parts (the sandbox, the policy engine, and the privacy router); the repository now documents four, adding a Gateway as the control-plane API that coordinates sandbox lifecycle and acts as the authentication boundary.^[9][23] OpenShell is Apache 2.0, written mostly in Rust, self-labelled alpha, and moving quickly: the repository was created on February 24, 2026 and released v0.0.92 on July 27, 2026.^[23] Detail on the sandbox internals, the policy language, the proposal-and-approval loop, and independent criticism belongs to the runtime's own article; see NVIDIA OpenShell as the main article, and the DGX Station for Windows page for how it is surfaced on that platform.

Omniverse libraries and MCP

Two extension points connect NemoClaw agents to the outside world. At SIGGRAPH 2026, NVIDIA added Omniverse libraries (ovrtx for RTX camera, lidar, radar, and sensor simulation; ovphysx for GPU-accelerated physics; and CAD-to-SimReady for converting CAD data into simulation-ready assets) that let a NemoClaw agent call physical AI capabilities, not just text tools.^[6] Separately, NemoClaw's documentation exposes a Model Context Protocol server so AI clients such as Claude Code and Cursor can connect to it, and NVIDIA's broader Agent Toolkit and Omniverse work uses MCP to plug agents into creative and engineering applications.^[6][7]

That MCP integration is worth distinguishing from a similarly worded claim about the runtime. NemoClaw ships packaged agent skills under .agents/skills/ that Cursor, Claude Code and other coding assistants can discover, and it publishes a documentation MCP server those assistants can query while helping a user install or operate a sandbox.^[21] That is Cursor consuming NemoClaw's documentation and skills. It is not the same as Cursor being a supported sandboxed agent: OpenShell's own supported-agent table lists Claude Code, OpenCode, Codex, GitHub Copilot CLI, OpenClaw and Hermes Agent (the last two sourced through NemoClaw), Ollama and Pi, and does not include Cursor, which appears only as a remote-editor option on openshell sandbox create --editor.^[23][28] Sandboxed agents run inside NemoClaw; Cursor talks to it from outside.

The LangChain Deep Agents blueprint and the cost claim

The flagship blueprint bundles three things: LangChain's Deep Agents Code harness, NVIDIA Nemotron 3 Ultra, and the NVIDIA OpenShell runtime, so a team can tune agents for its own workloads, run them securely, and optimize for quality, cost, and speed.^[3][4] The pitch is not that a bigger model wins, but that "agent performance improves when the model, harness, evals, and runtime are tuned together."^[3]

The attention-grabbing number is the cost claim. NVIDIA and LangChain say the blueprint delivers "benchmark-leading performance and more than 10x lower inference costs," a figure widely repeated in launch coverage.^[3][4][17] This is a vendor benchmark, and it is worth stating its exact basis rather than repeating the headline. In LangChain's own agent-evaluation suite, Nemotron 3 Ultra driven by Deep Agents reached an aggregate score of 0.86 at an inference cost of $4.48, while the next best-scoring model cost $43.48 for the same evaluation, roughly a 10x cost gap on that specific test.^[3][4] The figure therefore reflects one company-run benchmark on one eval suite, not an independently reproduced result; readers should treat "more than 10x lower inference costs" as a claim measured under conditions NVIDIA and LangChain chose. The blueprint was made available for enterprises to run against their own workloads on day one, which is how the companies invite verification.^[3][4]

The Hermes Agent research deployment

The most detailed public account of a working NemoClaw agent is a tutorial NVIDIA published on June 2, 2026 (updated June 11) by solutions architects Sam Pastoriza, Sean Lopp and Matthew Penn, building a product-research agent on the Hermes harness rather than OpenClaw or Deep Agents.^[8] The full example ships in the separate nemoclaw-community repository, created on May 6, 2026 and also Apache 2.0.^[8][25]

The agent answers research questions across two private channels (Slack and Outlook) and two public sources (GitHub issues and the NVIDIA developer forums), producing things like a daily digest of the issues that matter most on a repository. NVIDIA is explicit that the reporting is the demo and the security model is the point: what generalizes is "the pattern of safely mixing public and private data in a self-improving agent," which it lists as applying equally to sales research, customer support, engineering triage, competitive analysis, and internal knowledge discovery.^[8]

Model, harness, runtime

The tutorial states the three-layer division more crisply than any other NVIDIA source.

Scroll sideways for more →

The default model is nvidia/nemotron-3-super-120b-a12b served from build.nvidia.com, and NVIDIA notes that Hermes runs unchanged against the same model self-hosted on NIM or on vLLM "when traffic must stay on-prem."^[8] The prerequisites are modest for an enterprise agent: a host with a running Docker daemon (the example targets Ubuntu 24.04), an inference key, a GitHub token, and credentials for at least one messaging channel, either an Outlook tenant with a registered Azure app or a Slack workspace with a Slack app.^[8]

Credentials the agent never sees

Two properties of the deployment do the security work, and neither is implemented in the agent.

First, credential brokering. The Hermes process never holds the Slack or Outlook tokens; authentication is attached by OpenShell as requests leave the sandbox proxy.^[8] A prompt-injected agent cannot exfiltrate a secret it was never given.

Second, an asymmetry between private and public data. The agent reads sensitive Outlook and Slack content but is barred from the public internet entirely. Public community data does not arrive over the network at all: a separate host-side ETL job retrieves GitHub and NVIDIA forum content and stores it where the agent has read-only access, a design NVIDIA says "ensures that even a compromised agent can not post data to external sites."^[8] Read-only mirroring turns a two-way integration into a one-way one, which removes the channel an attacker would use rather than trying to detect its misuse.

Network policy as code

NVIDIA states the principle in one line: "The network policy is code, not a prompt." A policy.yaml file "declares every allowed destination, port, HTTP verb, and binary," so the inference authorization for this deployment reads:^[8]

network_policies:
  nvidia:
    endpoints:
      - host: integrate.api.nvidia.com
        port: 443
        rules:
          - allow: { method: POST, path: /v1/chat/completions }
          - allow: { method: POST, path: /v1/embeddings }
          - allow: { method: GET,  path: /v1/models }
    binaries:
      - path: /usr/local/bin/hermes
      - path: /usr/bin/python3

The allowlist is narrow in three directions at once: one host, three specific method-and-path pairs, and two named binaries. Anything else, including a different path on the same allowed host, fails. If the agent tries to reach a host that is not on the allowlist, the proxy returns 403 Forbidden and Hermes treats the refusal as an ordinary tool error rather than a crash, which lets the agent report the block and move on.^[8]

Observability

An agent that picks its own skills, tools and arguments is hard to debug from its output alone. The deployed agent records traces in Agent Trajectory Format (ATIF), and the sandbox image includes NVIDIA NeMo Relay by default so traces are captured without extra setup; scripts/download-traces.sh pulls them off the sandbox before it is torn down.^[8] Setting PHOENIX_COLLECTOR_ENDPOINT in the environment file streams the same traces live to an Arize Phoenix collector for interactive debugging.^[8]

Skills that persist

The self-improvement loop is deliberately unglamorous. A user asks for a daily update, dislikes the shape of the answer, specifies the exact format wanted (five issues, three discussions, each with number, title, state, URL and a one-line reason it matters, a bold header, and a two-to-three-sentence bottom line), then tells the agent to reuse that shape for everyone who asks. Hermes recognizes the pattern and writes a SKILL.md file to the filesystem with short YAML frontmatter (a name and description) and the format scaffolding as the body.^[8]

Surviving a rebuild is what makes the skill worth writing. NVIDIA walks through snapshot, tear-down, bring-up and restore, noting that production agents get rebuilt whenever code or configuration ships and that an agent whose skills do not survive has to be retaught every time. The snapshot captures skills, memories, sessions and scheduled jobs; a credential filter excludes .env and any file matching *token* or *secret* so the resulting tarball is safe to share.^[8] Asked from a fresh conversation after the restore, the agent returns the same format with only the underlying data changed, because "the format lives in the skill, not in conversational memory."^[8] The example also ships with five predefined skills that the gateway picks up automatically from agents/hermes/skills/, so operators can seed behavior before the agent learns any of its own.^[8]

What "self-evolving" means here

"Self-evolving" is doing specific and limited work in NVIDIA's marketing, and it is worth pinning down because the phrase invites much larger readings. It does not mean the model updates its own weights, and it does not mean the agent modifies its own policy. It means the agent writes durable artifacts, skill files and memories, into its own filesystem, and those artifacts change what it does on later runs.^[8][26] NVIDIA's Hermes coverage puts it plainly: "Hermes writes and refines its own skills. Every time the agent encounters a complex task or receives feedback, it saves its learnings as a skill so it can adapt and improve over time."^[26] The gain is that behavior taught once in a chat window applies to every later request from every user of that agent, without a code change, a redeploy, or a gateway restart.^[8]

That capability is also the security argument for putting enforcement outside the agent process. An agent that persists new instructions to disk is an agent whose future behavior is written by its inputs, and its inputs include untrusted material: GitHub issue text, forum posts, inbound email, messages from coworkers. Indirect prompt injection applies directly, and here it can persist, because a poisoned instruction that survives into a SKILL.md file keeps acting after the conversation that planted it has ended. Textual defenses do not help against this, since a system prompt, a refusal-trained model, and a tool description are all just more instructions to the same model that the attacker is manipulating. Guardrails that live inside the process being guarded fail with it.

The out-of-process answer is what NemoClaw's runtime provides, and the Hermes deployment shows the shape of it concretely. If the tokens live in OpenShell rather than in the agent, a compromised agent has no secret to leak. If the allowlist names hosts, ports, methods and binaries in a version-controlled file the agent cannot edit, a compromised agent's new "skill" for uploading data to an attacker's server returns 403 instead of succeeding. If public data reaches the agent only through a read-only host-side mirror, there is no outbound channel on that path to abuse in the first place.^[8][9] None of those three controls depends on the model behaving well, which is exactly why NVIDIA pairs an agent that rewrites itself with a runtime the agent cannot rewrite.

The Blender blueprint

At SIGGRAPH, NVIDIA showed what the Omniverse libraries make possible with a sample workflow for Blender, the open-source 3D creation tool. The Blender blueprint is a reference workflow "built in Blender with NVIDIA Omniverse libraries and NVIDIA NemoClaw" that brings agent-ready simulation into a tool artists already use: RTX sensor simulation, physics simulation, and asset validation become callable steps an agent can run while the human keeps creative control.^[6] The goal is to prepare 3D scenes and assets for physical AI and robotics pipelines (for example, generating synthetic data or a validated digital twin) without leaving the modeling application. It is a concrete demonstration that a NemoClaw agent is not limited to text and code: given the right libraries, it can act inside a simulator.^[2][6]

How it fits with the Agent Toolkit and DGX Station

NemoClaw is easy to confuse with the NVIDIA NeMo Agent Toolkit (marketed in 2026 as the "NVIDIA Agent Toolkit"), so the distinction is worth stating precisely from the sources. The Agent Toolkit is the broader framework: NVIDIA describes it as "the full deployment stack: models, tools, evaluation, and runtimes for building, testing, and optimizing long-running agents," and OpenShell is documented as "part of the NVIDIA Agent Toolkit."^[9] NemoClaw is the opinionated blueprint that assembles those parts: its product page states that "each blueprint includes NVIDIA Agent Toolkit components such as NVIDIA OpenShell for runtime policy controls," alongside model routing, skill execution, state, and observability.^[1] In short, the Agent Toolkit is the framework and parts catalog; NemoClaw is a reference build that packages those parts with a harness and a Nemotron model into something you can run today. At SIGGRAPH 2026 the two were announced together under the headline "NVIDIA Agent Toolkit Expands With New Omniverse Libraries," with NemoClaw presented as the agent-building on-ramp and the Omniverse libraries as the new capability the toolkit gained.^[6]

On hardware, NemoClaw targets NVIDIA's local-AI lineup: the DGX Station (a GB300 Grace Blackwell Ultra desktop) as the flagship for running the 550B Nemotron 3 Ultra locally, plus DGX Spark, RTX PRO workstations, and GeForce RTX PCs for smaller footprints.^[1][16] The recurring theme in NVIDIA's messaging is "frontier agents without API rent": keep the model, the data, and the runtime on a machine you own.^[1] NVIDIA's engineering-side support matrix is stricter than the marketing lineup, and the gap is documented below.

Supported platforms, agents, and integrations

NemoClaw maintains an unusually candid support matrix, generated from a checked-in ci/platform-matrix.json and reviewed through CODEOWNERS. NVIDIA states that any documentation, demo, blog post, sales conversation, or support reply describing a NemoClaw capability should match it, and that a row marked "Tested with limitations" must not be claimed as a "Tested" row.^[21] The matrix is the best correction available to NemoClaw's own marketing copy, and the following reflects it as of July 31, 2026.

Scroll sideways for more →

Two entries deserve emphasis because they contradict the marketing. RTX PCs and RTX PRO workstations appear in NVIDIA's product page and press release as NemoClaw targets, but the support matrix classifies that host class as Deferred: the managed vLLM path is available only behind an experimental flag, and the end-to-end onboarding path on that hardware was not yet validated in CI as of July 2026.^[1][19][21] And Outlook, the integration at the center of the Hermes research tutorial, is not in the supported messaging table at all; it comes from an example bridge in the community repository rather than from the core blueprint.^[8][21][25]

Availability

The NemoClaw for LangChain Deep Agents blueprint was released on July 8, 2026 and described as available immediately for enterprises to evaluate against their own workloads.^[3][4] NemoClaw itself is distributed through NVIDIA's open-source channels: the product page at nvidia.com/en-us/ai/nemoclaw, guided playbooks at build.nvidia.com/station, reference documentation at docs.nvidia.com/nemoclaw, and GitHub repositories.^[1][7][13] The build.nvidia.com "Run NemoClaw on DGX Station" playbook walks through installing Docker, Node.js, Ollama, the OpenShell CLI, and the NemoClaw plugin, then running the onboard wizard to create an OpenShell sandbox with inference routed to a local Nemotron model and a browser-based chat dashboard.^[13] Installation is designed to be a single command that sets up OpenClaw, the OpenShell runtime, and an open Nemotron model with hardened defaults for networking, data access, and security.^[10] NVIDIA also ran a hands-on "Build a Claw" experience showing an autonomous agent built with Nemotron and OpenClaw.^[15]

Two repositories carry the code. NVIDIA/NemoClaw is the reference stack itself, Apache 2.0 and mostly TypeScript, with an interactive terminal installer and a documented alternative of handing a starter prompt to a local coding agent that then installs NemoClaw with the user.^[20] NVIDIA/nemoclaw-community holds worked examples, including the Hermes research deployment described above.^[25] NVIDIA describes NemoClaw as "available in early preview starting March 16, 2026" and tracks changes across three CLIs, NemoClaw, NemoHermes and NemoDeepAgents, in one set of release notes.^[29] Cadence is high: the repository was still taking pushes on July 31, 2026, and its tag series had reached v0.0.98.^[20]

Significance

NemoClaw matters less as a new model than as a stance on how agents should be built and run. Three ideas stand out. First, it argues that agent quality comes from tuning the model, harness, evaluation, and runtime together, and it ships that whole system as a blueprint rather than leaving teams to integrate four projects themselves.^[3][4] Second, it is a local-first, open-stack answer to metered cloud APIs: an open Nemotron model plus OpenShell plus the enterprise's own data, running on a workstation the company owns, with the "more than 10x lower inference costs" claim as the economic argument.^[1][4] Third, by treating the harness as a swappable slot filled by community and partner projects (OpenClaw, Hermes, Deep Agents) rather than an NVIDIA-only framework, NVIDIA positions itself as the infrastructure layer (model, runtime, hardware, packaging) under whichever agent framework a team prefers.^[7][10] The SIGGRAPH extension into Omniverse and Blender then stretches the same agents from text and code toward simulation and physical AI, which is where NVIDIA wants the next wave of agents to operate.^[2][6]

There is a fourth idea, less advertised but visible in the Hermes deployment: NemoClaw is a bet that the way to make an agent trustworthy is not to make the model safer but to shrink what the model is allowed to touch. Credentials it never receives, hosts it cannot reach, and public data it can only read are all controls that hold whether or not the model is behaving.^[8][9]

Limitations and caveats

• New and fast-moving. NemoClaw was announced in March 2026 and the Deep Agents blueprint in July 2026. Interfaces, defaults, and even the set of supported harnesses are likely to shift; treat exact command lines and version-specific details as provisional.^[3][7][19]
• Alpha software by NVIDIA's own label. NemoClaw is stage "alpha" and label "early preview" in its support matrix, with no production SLA and best-effort maintainer response.^[20][21] OpenShell describes itself the same way, as proof-of-life "single-player mode" for one developer, one environment and one gateway.^[23]
• Version drift is real. The June 2026 Hermes tutorial pins OPENSHELL_VERSION=v0.0.38, and NemoClaw's most recent published compatibility review pins a stable OpenShell of v0.0.72, reviewed on June 29 and 30, 2026, while OpenShell itself shipped v0.0.92 on July 27, 2026.^[8][23][30] Command lines and pinned versions in any NemoClaw blog post age within weeks. A concrete example: NVIDIA's March 2026 OpenShell launch post advertised openshell sandbox create --remote spark --from openclaw, but the current CLI reference has no --remote flag on sandbox create; --remote <USER@HOST> belongs to openshell gateway add.^[9][28]
• Vendor benchmarks. "Benchmark-leading performance" and "more than 10x lower inference costs" are claims from NVIDIA and LangChain, measured on LangChain's own eval suite (0.86 aggregate at $4.48 versus $43.48 for the next model).^[3][4] They had not been independently reproduced at launch. Real-world cost and quality depend on the workload, the model size chosen, and how much tuning a team does.
• Not a turnkey product for every model. The headline economics rest on running a large open model (Nemotron 3 Ultra, 550B) locally, which in practice means DGX Station GB300-class hardware; smaller machines run smaller Nemotron variants with different performance.^[2][8]
• Marketing hardware exceeds validated hardware. RTX PCs and RTX PRO workstations are named as NemoClaw targets on the product page and in the GTC press release, but NVIDIA's engineering support matrix marks that host class Deferred, with no dedicated CI job, as of July 2026.^[1][19][21]
• Overlapping names. NemoClaw, OpenClaw, OpenShell, the NeMo Agent Toolkit, and Nemotron are distinct layers that are easy to conflate. OpenClaw in particular is a third-party open-source project, not an NVIDIA harness, despite the shared "Claw" branding.^[7][10] The correct casing throughout is NemoClaw, matching the repository and NVIDIA's own materials, not "NeMoClaw" by analogy with NeMo.^[19][20]
• Scope of MCP support. Public documentation confirms an MCP server for connecting AI clients (Claude Code, Cursor) to NemoClaw; broader MCP client and tool integration is part of NVIDIA's wider Agent Toolkit and Omniverse story rather than fully detailed on the NemoClaw page itself.^[6][7] That MCP and skills integration is also the only sense in which Cursor is supported: it is absent from OpenShell's supported-agent table, where it appears only as a remote-editor flag.^[23][28]
• Sandbox hardening has a documented soft spot. NemoClaw drops dangerous Linux capabilities by default, but the fail-closed gate is opt-in via NEMOCLAW_REQUIRE_CAP_DROP=1; the default is warn-and-continue so hosts lacking CAP_SETPCAP still boot, which leaves dangerous capabilities in the bounding set on some hosts. NVIDIA tracks this as an active issue.^[21]

ELI5

Imagine you want a robot assistant that runs by itself all day, checking things and doing tasks without you typing every command. To build one you normally need three hard-to-fit pieces: a brain (the AI model), a control loop that decides what to do next and remembers things (the harness), and a safety cage so it cannot touch files or send data it should not (the runtime). NemoClaw is a set of ready-made kits from NVIDIA that snap those three pieces together for you. It brings a strong open brain (a Nemotron model), lets you pick one of three control loops (OpenClaw, Hermes, or LangChain's Deep Agents), and locks it inside a safety cage called OpenShell, all set up with one command on an NVIDIA computer you own. NVIDIA says its kit is cheaper to run than paying a cloud service by the message, and at SIGGRAPH it showed the same assistant learning to work inside 3D and physics tools like Blender.

The safety cage matters more than it sounds, because these assistants are designed to learn. In NVIDIA's own research example, you teach the assistant your preferred report format once in chat and it writes itself a note so it uses that format forever, even after the computer is wiped and rebuilt. An assistant that can rewrite its own instructions is an assistant a stranger might be able to rewrite too, by hiding a message in an email or a public forum post it reads. So the passwords are kept outside the cage where the assistant cannot see them, and a written list decides which websites it may talk to. If it tries anything else, the door simply does not open.

See also

• NVIDIA NeMo Agent Toolkit
• NVIDIA OpenShell
• OpenClaw
• Nemotron 3
• NVIDIA DGX Station
• Deep Agents
• LangChain
• Nous Research
• Model Context Protocol
• NVIDIA Omniverse

References

1. ^NVIDIA, "NVIDIA NemoClaw" (product page), nvidia.com/en-us/ai/nemoclaw/ (accessed July 2026). nvidia.com/...nemoclaw
2. ^NVIDIA, "At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI," NVIDIA Blog, July 20, 2026. blogs.nvidia.com/...siggraph-news-2026
3. ^LangChain, "LangChain and NVIDIA Launch the NemoClaw Deep Agents Blueprint," LangChain Blog, July 8, 2026. langchain.com/...he-nemoclaw-deep-agents-blueprint
4. ^LangChain / NVIDIA (via PR Newswire), "LangChain and NVIDIA Launch NemoClaw Deep Agents Blueprint for Enterprise Agents," July 8, 2026. prnewswire.com/...-for-enterprise-agents-302820446
5. ^AIwire (HPCwire), "LangChain and NVIDIA Launch NemoClaw Deep Agents Blueprint for Enterprise Agents," July 8, 2026. hpcwire.com/...nts-blueprint-for-enterprise-agents
6. ^NVIDIA Newsroom, "NVIDIA Agent Toolkit Expands With New Omniverse Libraries, Putting AI Agents to Work Building Simulation-Ready Worlds," July 20, 2026. nvidianews.nvidia.com/...g-simulation-ready-worlds
7. ^NVIDIA, "NemoClaw Quickstart with OpenClaw," NVIDIA NemoClaw User Guide (docs), accessed July 2026. docs.nvidia.com/...quickstart
8. ^NVIDIA, "Deploy Self-Evolving Agents for Faster, More Secure Research with a Hermes Agent and NVIDIA NemoClaw," NVIDIA Technical Blog, 2026. developer.nvidia.com/...-agent-and-nvidia-nemoclaw
9. ^NVIDIA, "Run Autonomous, Self-Evolving Agents More Safely with NVIDIA OpenShell," NVIDIA Technical Blog, 2026. developer.nvidia.com/...fely-with-nvidia-openshell
10. ^NVIDIA, "Nemotron Labs: What OpenClaw Agents Mean for Every Organization," NVIDIA Blog, April 30, 2026. blogs.nvidia.com/...ts-mean-for-every-organization
11. ^NVIDIA, "NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents," NVIDIA Technical Blog, 2026. developer.nvidia.com/...ng-for-long-running-agents
12. ^NVIDIA Research, "NVIDIA Nemotron 3 Ultra," Nemotron Labs, accessed July 2026. research.nvidia.com/...Nemotron-3-Ultra
13. ^NVIDIA, "Run NemoClaw on DGX Station" (playbook), build.nvidia.com/station, accessed July 2026. build.nvidia.com/...overview
14. ^NVIDIA, "Build a More Secure, Always-On Local AI Agent with OpenClaw and NVIDIA NemoClaw," NVIDIA Technical Blog, 2026. developer.nvidia.com/...idia-nemoclaw-and-openclaw
15. ^NVIDIA, "Build an Autonomous AI Agent with Nemotron and OpenClaw at GTC," nvidia.com/en-us/ai/build-a-claw/ (accessed July 2026). nvidia.com/...build-a-claw
16. ^Wccftech, "NVIDIA Brings Local AI Agents To Its Most Powerful Workstation PC, The DGX Station, With the NVIDIA Agent Toolkit and Omniverse," July 2026. wccftech.com/...-dgx-station-pc-with-agent-toolkit
17. ^TechTimes, "Open Agent Stack Cuts Enterprise AI Cost 10x: NemoClaw Blueprint Ships," July 9, 2026. techtimes.com/...cost-10x-nemoclaw-blueprint-ships
18. ^LangChain, "Deep Agents," langchain.com/deep-agents (accessed July 2026). langchain.com/deep-agents
19. ^NVIDIA Newsroom, "NVIDIA Announces NemoClaw for the OpenClaw Community," March 16, 2026. nvidianews.nvidia.com/...nvidia-announces-nemoclaw
20. ^NVIDIA, "NVIDIA NemoClaw: Reference Stack for Sandboxed AI Agents in OpenShell" (README and repository metadata), GitHub, accessed July 31, 2026. github.com/...NemoClaw
21. ^NVIDIA, "Platform Support and Launch Claims," NVIDIA NemoClaw documentation, accessed July 31, 2026. docs.nvidia.com/...platform-support
22. ^NVIDIA, "Ecosystem," NVIDIA NemoClaw documentation, accessed July 31, 2026. docs.nvidia.com/...ecosystem
23. ^NVIDIA, "OpenShell: the safe, private runtime for autonomous AI agents" (README and repository metadata), GitHub, accessed July 31, 2026. github.com/...OpenShell
24. ^NVIDIA, "NVIDIA-Nemotron-3-Super-120B-A12B-BF16" (model card), Hugging Face, accessed July 31, 2026. huggingface.co/...-Nemotron-3-Super-120B-A12B-BF16
25. ^NVIDIA, "nemoclaw-community: Our repository of community driven examples, showcases, and integrations," GitHub, accessed July 31, 2026. github.com/...nemoclaw-community
26. ^NVIDIA, "Hermes Unlocks Self-Improving AI Agents, Powered by NVIDIA RTX PCs and DGX Spark," NVIDIA Blog, May 13, 2026. blogs.nvidia.com/...-garage-hermes-agent-dgx-spark
27. ^Nous Research, "hermes-agent: The agent that grows with you" (repository), GitHub, accessed July 31, 2026. github.com/...hermes-agent
28. ^NVIDIA, "OpenShell CLI Reference" (.agents/skills/openshell-cli/cli-reference.md), NVIDIA/OpenShell, GitHub, accessed July 31, 2026. github.com/...cli-reference.md
29. ^NVIDIA, "Release Notes," NVIDIA NemoClaw documentation, accessed July 31, 2026. docs.nvidia.com/...release-notes
30. ^NVIDIA, "OpenShell 0.0.72 Compatibility Review," NVIDIA NemoClaw documentation, reviewed June 29 to 30, 2026, accessed July 31, 2026. docs.nvidia.com/...ell-0.0.72-compatibility-review