At GTC Berlin on Oct 21, NVIDIA announced the OpenShell Reference Architecture: building on the open-source OpenShell secure runtime, it pairs Vera Rubin for inference, Vera CPUs for agent harnesses, and BlueField-4 DPUs running NVIDIA Sentry for in-silicon policy enforcement, so enterprises can build, contain, observe and govern agents on-prem and at the edge. NVIDIA says it is available now; partners named include Anthropic, OpenAI, Meta, Microsoft and ServiceNow. The software is Apache 2.0.
Key Takeaways
- ✓Full stack: Vera Rubin for model inference, Vera CPUs for agent harnesses/tool use, BlueField-4 + Sentry for in-silicon policy and quarantine (vendor claims)
- ✓Vendor claims: Vera Rubin NVL72 about 30x throughput per megawatt on agentic workloads; Sentry can quarantine a boundary-breaking agent in milliseconds and boost agent capacity by up to 20%
- ✓Runtime: deny-by-default, enforcement outside the agent process; sandboxes have no direct network, syscalls are filtered in-kernel, credentials are injected only onto allowed requests
- ✓Stated agent paths include Claude Code, Codex, GitHub Copilot CLI, OpenCode, OpenClaw, LangChain Deep Agents; deploy with Docker, Podman or Kubernetes via Helm
- ✓Start: curl install script then openshell sandbox create; docs at docs.nvidia.com/openshell; GitHub NVIDIA/OpenShell (Apache 2.0) and PyPI package openshell
Key Decision Metrics at a Glance
Heavy Claude Code use: compare subscription limits and API bills
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the pain point
Enterprises want agents that read files and call APIs without open host or secret access. Containers isolate processes but do not stop prompt-bypassable policy or long-lived egress credentials. On Oct 21, 2026 at GTC Berlin, NVIDIA announced the OpenShell Reference Architecture so policy is enforced outside the agent process.
Architecture and mechanism
Built on open-source OpenShell: Vera Rubin for inference, Vera CPUs for harnesses and tools, BlueField-4 Sentry on DOCA for out-of-band monitoring and in-silicon enforcement. Sandboxes, a formal Policy Prover, gateway and supervisor; no direct network, filtered syscalls, credentials only on approved endpoints. Runs under Claude Code, Codex, OpenCode, OpenClaw and similar; it is not another agent framework.
Benchmarks
Figures are vendor-reported; no independent agent-safety leaderboard in the launch materials. NVIDIA claims about 30x throughput per megawatt on Vera Rubin NVL72 for agentic work, about 2x x86 on Vera CPUs, and millisecond quarantine plus up to 20% more agent capacity via Sentry. Re-measure on your own stack.
Getting started
Available now per NVIDIA. GitHub NVIDIA/OpenShell (Apache 2.0), docs.nvidia.com/openshell, PyPI openshell. Install via curl, then openshell sandbox create --name demo; fleets can Helm-deploy the gateway with NetworkPolicy. ServiceNow announced Arc GA with Control Tower policies enforced by OpenShell/Sentry. Start with the open runtime locally; silicon enforcement needs BlueField-4 and Vera/Rubin planning.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.