Running AI agents locally has become technically feasible — but making them practical for business has been a different challenge. NVIDIA's NemoClaw, released in early preview at GTC 2026, combines the OpenClaw always-on assistant with a fully sandboxed local inference stack via Ollama. The result: an AI agent that stays entirely on your hardware and never phones home.
For European businesses navigating GDPR and the EU AI Act, that architectural guarantee carries more weight than any data processing agreement with a cloud vendor.
What NemoClaw Actually Is
NemoClaw is not a language model. It is an open-source reference stack that layers three components:
- OpenClaw — the always-on AI assistant that handles conversations, tool dispatch, and multi-step task execution.
- OpenShell — NVIDIA's sandboxed runtime that enforces OS-level network and filesystem isolation.
- Ollama — the local model runner that keeps all inference on-device.
The OpenShell layer is what distinguishes NemoClaw from a plain Ollama wrapper. Network egress is blocked by default and must be explicitly enabled per policy preset. Filesystem access is scoped. Credentials are isolated from the runtime. These are not advanced configuration options — they are defaults.
Setup: One Script, One Wizard
A single installer script (nemoclaw.sh) handles all dependencies: Node.js, OpenShell, and the NemoClaw CLI. The post-install wizard walks through four choices:
- Model selection: lists installed Ollama models or offers starter suggestions if none are present
- Web search: optionally enables Brave Search for live-web queries
- Messaging channel: Telegram, Discord, or Slack
- Network preset: from fully offline to selective outbound access
Everything is stored in a local JSON config — version-controllable, auditable, no vendor lock-in.
Nemotron 3 Super as the Reference Model
NVIDIA's own recommendation for NemoClaw is Nemotron 3 Super. At 120B total parameters with only 12B active per forward pass (Mixture-of-Experts architecture), it offers a practical inference profile for high-end hardware. According to NVIDIA's published benchmarks, Nemotron 3 Super scores 85.6% on PinchBench — a benchmark designed specifically for tool-calling, multi-step planning, and agentic task execution — ranking first among open-weight models at the time of release.
Hardware requirements are substantial: the model needs approximately 76–80 GB of VRAM or unified memory at Q4KM quantization, and around 87 GB of disk space. This puts it in the territory of a well-equipped GPU workstation or server. NVFP4 optimization requires CUDA; Apple Silicon users will need community quantization ports, as native Apple Silicon optimization is not documented at time of writing.
For teams running Mac Studio M3 or M4 Ultra (64–192 GB unified memory), smaller models like Qwen3:35b support tool-calling and work cleanly with NemoClaw — the setup wizard will suggest them when available GPU memory is below the threshold for Nemotron 3 Super.
Business Use Cases
Document Intelligence Without Cloud Exposure
A NemoClaw agent with read access to a local folder of contracts, SOPs, or supplier records can answer questions, draft summaries, and flag inconsistencies. Nothing leaves the building. This is particularly relevant for law firms, accounting practices, healthcare operators, or any business subject to professional secrecy obligations.
Team-Facing Assistant via Messaging
With Slack or Telegram integration, any team member can query the agent about internal processes, project status, or technical specifications. The agent responds through the messaging channel; all processing stays on the local server.
Code Review and Development Support
Paired with a code-capable model — Llama 3.3, Qwen3-Coder, or DeepSeek-V3 at appropriate quantization — NemoClaw provides a persistent code assistant accessible to the entire development team, without API keys, without cloud subscriptions, and without usage-based billing.
GDPR and EU AI Act Implications
Under GDPR, routing employee or customer data through a third-party cloud AI service triggers documentation obligations, potential cross-border transfer restrictions, and Article 28 processor agreements. A fully local stack removes these data flows at the architecture level, significantly shrinking compliance surface area.
The EU AI Act adds further obligations for deployers of AI systems in high-risk categories (hiring, legal proceedings, credit assessment). Under Article 26, deployers must ensure transparency and maintain meaningful oversight of their systems. A sandboxed local agent running on auditable JSON configuration is considerably easier to document and supervise than a black-box cloud API.
Based on our reading of current guidance, local-first architectures like NemoClaw offer a structurally stronger compliance baseline — though the specifics of your implementation and use case always warrant independent legal review.
Current Maturity
NemoClaw is in early preview as of this writing. The Ollama integration is marked experimental in the official documentation. For business-critical production workflows, waiting for a stable release — or working with experienced engineers during rollout — is advisable.
For pilot projects, internal toolchains, and architecture evaluation, the stack is functional today. The underlying components (Ollama, local model, OpenShell sandbox, messaging integration) are each independently stable.
Getting Started with Freshlab
Freshlab helps SMBs across Europe evaluate and deploy local AI infrastructure that fits their data sovereignty requirements. Our pilot project programme provides a structured path from initial evaluation to production rollout, covering hardware sizing, model selection, and privacy policy configuration.
The kAIra Toolkit integrates natively with local Ollama deployments, adding pre-built agent components for document processing, internal search, and team communication — complementing what NemoClaw provides at the infrastructure layer.
Ready to explore what local AI agents could do for your business? Get in touch.