AI Defense Stack

Agent & MCP security

NVIDIA OpenShell: open-source runtime for sandboxed AI agents

OpenShell is an open-source runtime for sandboxed AI agents from NVIDIA. It is released under the Apache-2.0 license and is offered as: CLI, Self-hosted, Cloud-native.

  • Last reviewed
Visit OpenShell (opens in a new tab) Documentation (opens in a new tab) Source code (opens in a new tab)
Open sourceApache-2.0 Layer 7: Agent and tool access

Best for

Teams that run coding agents or other autonomous agents and want per-sandbox file, process and network policy with credentials kept outside the agent

Consider instead if

  • You want a Kubernetes resource that delegates isolation to a runtime such as gVisor or Kata Containers. See Kubernetes Agent Sandbox.

What OpenShell does

OpenShell is an open-source (Apache-2.0) runtime from NVIDIA for running autonomous AI agents inside sandboxes. The docs say it uses kernel-level isolation, with Landlock restricting file access and seccomp limiting system calls, and declarative YAML policies for filesystem, process, network and provider credentials. Outbound traffic is allowed only to approved destinations, and credentials are added only to requests bound for approved endpoints, so the agent never sees them. The docs list Claude Code, OpenCode, Codex and GitHub Copilot CLI as supported agents. NVIDIA's September 28, 2026 release presents OpenShell together with Sentry, an out-of-band watchdog on BlueField-4 DPUs, as its Open Agent Safety Platform.

This summary is our own wording of the sourced facts on this page.

Key facts

Each value links to the page that states it, with the date we checked it.

Vendor
NVIDIA Source, checked October 21, 2026: nvidianews.nvidia.com (opens in a new tab)
Status
Active Source, checked October 21, 2026: api.github.com (opens in a new tab)
Open source
Yes: Apache-2.0 Source, checked October 21, 2026: api.github.com (opens in a new tab)
Repository
github.com Source, checked October 21, 2026: api.github.com (opens in a new tab)
GitHub stars
17,031 as of October 21, 2026 Source, checked October 21, 2026: api.github.com (opens in a new tab)
Latest push
October 22, 2026 Source, checked October 21, 2026: api.github.com (opens in a new tab)
Latest release
October 20, 2026 Source, checked October 21, 2026: api.github.com (opens in a new tab)
Deployment
CLI, Self-hosted, Cloud-native Source, checked October 21, 2026: github.com (opens in a new tab)
Pricing model
Free Source, checked October 21, 2026: api.github.com (opens in a new tab)
Website
docs.nvidia.com Source, checked October 21, 2026: docs.nvidia.com (opens in a new tab)
Documentation
docs.nvidia.com Source, checked October 21, 2026: docs.nvidia.com (opens in a new tab)
API reference
Not documented in the sources we checked
Quickstart
Not documented in the sources we checked
Security page
github.com Source, checked October 21, 2026: github.com (opens in a new tab)
Attestations or disclosure
Coordinated vulnerability disclosure through NVIDIA PSIRT, with no bug bounty program Source, checked October 21, 2026: github.com (opens in a new tab)

Capabilities

As documented by the vendor or project. The tag is our label from a controlled vocabulary.

Integrations

Where OpenShell fits in our AI security layers

Our analysis

Our editorial model, not a standard. The primary layer follows the category, except for tools whose main documented job is detection and response; other layers rest on a capability the sources document. How the model works.

AI security layers that OpenShell covers, with the role, how each applies and what it rests on
LayerRoleHow it appliesBased on
L7 Agent access Agent and tool access Primary Limit which tools, MCP servers, data and credentials an agent can use, and require approval for risky actions. Follows its category, Agent & MCP security

OWASP LLM Top 10 mapping

Our analysis

The risk list is OWASP's (Top 10 for LLM Applications, 2025). The mapping is our analysis, and each row rests on a capability the sources document.

OWASP LLM risks mapped to OpenShell, with the capabilities each mapping rests on
RiskHow it appliesBased on
LLM02 Sensitive Information Disclosure Credentials reach only approved endpoints, so the agent never holds a real credential that it could leak. Adds credentials only to requests bound for approved endpoints, so the agent never sees the real credential
LLM06 Excessive Agency Sandbox isolation and YAML policies limit which files, processes and network destinations an agent can use. Runs each agent in an isolated sandbox, with Landlock limiting file access and seccomp limiting system calls; Declarative YAML policies cover filesystem, process, network and provider credentials, and outbound traffic is allowed only to approved destinations

Not mapped in our analysis: LLM01, LLM05, LLM07. These are risks we expect tools in this category to cover; the sources we checked do not document a matching capability.

Alternatives and related tools

More in Agent & MCP security, ranked within the category by Coverage Score.

Cisco DefenseClaw

Cisco

Open sourceApache-2.0
  • CLI
  • Self-hosted

Scans skills and MCP servers, inspects agent tool calls and keeps an audit trail around AI coding agents.

Best for: Security teams that want scanning, guardrails and an audit trail around the coding agents developers already run, starting in observe mode before anything is blocked

Coverage Score 76 out of 100
  • LLM01 Prompt Injection
  • LLM02 Sensitive Information Disclosure
  • LLM06 Excessive Agency

Reviewed

Docker MCP Gateway

Docker

Open sourceMIT
  • CLI
  • Self-hosted

Proxies AI clients to MCP servers it runs as isolated Docker containers, with tool filters and call logs.

Best for: Developers and platform teams that already use Docker and want MCP servers to run in containers behind one gateway with per-server tool selection

Coverage Score 72 out of 100
  • LLM02 Sensitive Information Disclosure
  • LLM06 Excessive Agency

Reviewed

agentgateway

agentgateway (a Series of LF Projects, LLC)

Open sourceApache-2.0
  • CLI
  • Self-hosted
  • Cloud-native

Open-source proxy that applies auth, CEL policy and guardrails to LLM, MCP and agent-to-agent traffic.

Best for: Platform teams that want one open-source gateway in front of MCP servers and LLM providers, with token validation, CEL rules per MCP method and pluggable guardrails

Coverage Score 69 out of 100
  • LLM02 Sensitive Information Disclosure
  • LLM06 Excessive Agency

Reviewed

OpenShell FAQ

Is OpenShell open source?

Yes. It is released under Apache License 2.0, and the code is on GitHub at NVIDIA/OpenShell.

Which agents does it support?

The docs list Claude Code, OpenCode, Codex and GitHub Copilot CLI as supported agents.

What is Sentry?

NVIDIA's press release of September 28, 2026 describes Sentry as an out-of-band watchdog that runs on BlueField-4 DPUs to monitor agent behavior. The release calls it the Sentry reference system design, the hardware part of the Open Agent Safety Platform next to the OpenShell software, and it is not part of the OpenShell repository.

Key facts you can cite

We have no commercial relationship with NVIDIA.

Sources & last reviewed

Last reviewed . We compile from the vendors' own public pages and repositories and we do not run the products. Source links open in a new tab. Found a mistake? Report a correction.