Overview

Swif's Agentic Security policies control and monitor AI agents, such as Claude Code, Codex CLI, Cursor, and OpenClaw, on managed macOS, Windows, and Linux devices. Use this page to find the right policy, check what it requires, and open its full article.

Before you start

Install the Swif Agentic Security agent first. These policies configure the agent; they don't install it. See Getting Started with Swif's Agentic Security for Enterprise.
Install OpenClaw for the OpenClaw policies. The three OpenClaw policies need the OpenClaw gateway on the device.
Install NVIDIA OpenShell for the OpenShell policies.
- The OpenShell Sandbox Policy needs OpenShell 0.1 or later on macOS or Linux.
- The Require OpenShell for Confidential Code Policy needs OpenShell, with the Sandbox Policy, only to block. Without them, it still logs agents that work on confidential code outside OpenShell.
- The OpenShell Inventory Policy works without OpenShell, and reports OpenShell not installed.
Open each AI agent at least once. Some policies, such as AI Agent Command Protection, need the agent's local files to exist on the device.
Create policies under Agentic Security. In the Swif Admin Console, go to Agentic Security > Policies and select Create new policy.

This covers all three OpenShell policies, since each needs OpenShell differently:

Policy	Needs OpenShell?
OpenShell Sandbox Policy	Yes. There’s nothing to apply the policy to without it.
Require OpenShell for Confidential Code Policy	Only to block. Logging works without it.
OpenShell Inventory Policy	No. It reports OpenShell not installed.

Available policies

Monitoring

Policy	What it does	Platforms	Device ownership	Learn more
Swif Agentic Security Logging Policy	Turns audit logging for supported AI-agent activity on or off. Logged actions appear in the device's Agentic Security view, rated Safe, Risky, or Danger.	macOS, Windows, Linux	Company-owned, BYOD	Article
Swif AI Agent Activity Monitoring Policy	The Swif AI Agent Activity Monitoring Policy records the sessions and tool activity of AI coding agents on managed macOS, Windows, and Linux devices, so admins can review what each agent did.	macOS, Windows, Linux (Claude Desktop and Warp: macOS and Windows only)	Company-owned, BYOD	Article
Swif OpenShell Inventory Policy	Collects an inventory of NVIDIA OpenShell for each local user on the device: installs and versions, gateways, running sandboxes with their AI agent and image digest, and the active policy hash. Use it to find OpenShell across your fleet and confirm sandboxes run the expected policy.	macOS, Linux (not Windows)	Company-owned, BYOD	Article
Swif Agent Inventory Policy	Finds autonomous AI agents installed on devices (Hermes and Grok Bot), and reports for each user: install and version, background service, running processes, risky launch flags, and a summary of security settings. Doesn’t install, change, or stop any agent.	macOS, Linux, Windows (Grok Bot: macOS only)	Company-owned, BYOD	Article

Blocking

Policy	What it does	Platforms	Device ownership	Learn more
Swif AI Agent Command Protection Policy	Blocks destructive commands, such as `rm -rf` and `git reset --hard`, before Claude Code, Codex CLI, or Cursor runs them. Supports an allowlist, and the agents can't turn it off themselves.	macOS (Apple silicon and Intel), Windows 11 (x64), Linux (Ubuntu 22.04+, Debian 12, Rocky Linux/RHEL 9)	Company-owned, BYOD	Article
Swif MCP Security Policy	Inventories the MCP servers that Claude Code, Cursor, Claude Desktop, and Codex are set up to use, and reports or blocks any that your allowlist or blocklist doesn’t allow. Blocking works in Claude Code and Cursor; Claude Desktop and Codex servers are reported only. Records only each server’s package or host, never its arguments, environment values, or tokens.	macOS, Linux, Windows	Company-owned, BYOD	Article

Sandbox controls

Policy	What it does	Platforms	Device ownership	Learn more
Swif OpenShell Sandbox Policy	Sets one Swif-managed policy for every NVIDIA OpenShell sandbox on the device, including sandboxes created later. Controls which filesystem paths, network destinations, and model endpoints sandboxed agents can reach, and blocks everything else. Users can't change it from inside a sandbox.	macOS, Linux (not Windows)	Company-owned, BYOD	Article
Require OpenShell for Confidential Code Policy	Keeps Claude Code and Codex inside OpenShell when they work on confidential folders or Git repositories. When an agent works on a confidential location outside a sandbox, the action is reported or blocked, and the agent is told to run inside OpenShell. Blocking requires the OpenShell Sandbox Policy on the same device.	macOS, Linux, Windows (Windows: report only)	Company-owned, BYOD	Article

OpenClaw controls

Policy	What it does	Platforms	Device ownership	Learn more
Swif OpenClaw Tools Access Policy	Controls which tools OpenClaw agents can use: a base tool profile, allowed and denied tools, rules by provider and sender, sandbox rules, elevated commands, agent-to-agent access, and which sessions an agent can see.	macOS, Windows, Linux	Company-owned, BYOD	Article
Swif OpenClaw Tools Runtime Policy	Controls how permitted tools run: command timeouts and background behavior, the `apply_patch` file-editing tool, loop detection, and limits on files passed to spawned sessions.	macOS, Windows, Linux	Company-owned, BYOD	Article
Swif OpenClaw Tools Experimental Policy	Turns OpenClaw's multi-step planning tool on or off. Off by default. Newer OpenClaw versions changed how this feature is configured, so test it with your OpenClaw version.	macOS, Windows, Linux	Company-owned, BYOD	Article

How the policies fit together

Monitoring policies record what agents do. Blocking, sandbox, and OpenClaw policies limit what they can do. Use both for complete coverage.

Goal	Use
See what AI agents did on a device	Logging Policy, AI Agent Activity Monitoring Policy
Stop agents from deleting files or Git history	AI Agent Command Protection Policy
Limit which tools an OpenClaw agent can use	OpenClaw Tools Access Policy
Limit how long OpenClaw commands run, or restrict file editing	OpenClaw Tools Runtime Policy
Turn OpenClaw's planning feature on or off	OpenClaw Tools Experimental Policy
Limit which files, hosts, and model APIs agents in OpenShell sandboxes can reach	OpenShell Sandbox Policy
Make sure agents work on confidential code only inside OpenShell	Require OpenShell for Confidential Code Policy, with the OpenShell Sandbox Policy
Find where OpenShell is installed and which sandboxes are running	OpenShell Inventory Policy
See which MCP servers AI agents use, and block unapproved ones	MCP Security Policy
Find autonomous AI agents, such as Hermes or Grok Bot, and how they’re set up	Agent Inventory Policy

A few things to keep in mind:

Monitoring doesn’t block anything. The Logging, AI Agent Activity Monitoring, OpenShell Inventory, and Agent Inventory policies record activity or report what’s installed, but don’t stop a risky action.
Access rules take priority in OpenClaw. The Runtime and Experimental policies can't make a tool available if the Tools Access Policy denies it.
Assign one policy of each type per device. If two policies of the same type conflict, the result can depend on which was applied last.
Removing a policy may not undo its settings. For the OpenClaw policies, deploy the settings you want to keep, confirm them, and only then remove the assignment.
The OpenShell Sandbox Policy blocks anything not listed. When it's enabled, sandboxed agents can reach only the hosts and model endpoints in the policy. List every destination agents need before you turn it on.
The Require OpenShell for Confidential Code Policy needs the Sandbox Policy to block. Without the OpenShell Sandbox Policy on the same device, matches are reported but not blocked, and the device is flagged. Start in log mode to see what would be blocked.
MCP Security blocking depends on the agent. In block mode, Claude Code doesn’t start servers that aren’t allowed, and Cursor refuses their tool calls. Claude Desktop and Codex servers are only reported. Start in log mode to see what would be blocked.

Privacy

The Logging and AI Agent Activity Monitoring policies can collect prompts, responses, commands, file paths, and tool results, and the OpenShell Inventory Policy reports which user runs each install and sandbox.

The MCP Security Policy records only which package or host each MCP server runs, never its arguments, environment values, or tokens.

The Agent Inventory Policy reports which agents each user has installed and how they’re set up.

Before you turn them on, especially on BYOD devices:

Tell users what's collected.
Limit who in your organization can view the data.
Set retention rules for it.

Each policy's article has more detail.

How to use this list

Find the policy you need in the tables above.
Check the platforms and requirements. Devices that don't meet them don't apply the policy.
Open the article for settings, examples, verification steps, and troubleshooting.
Test on a small pilot group before a wider rollout.

List of Agentic Security Policies