Set up Vardhold.
Install Vardhold on one Windows PC, connect the AI agents on it, and watch it make its first decisions. Allow about fifteen minutes.
On this page
New to Vardhold? Follow the four steps below. Rolling it out to a fleet? Go straight to Install and deploy.
Vardhold is in private pilot. Installers are shared with pilot customers directly, signed by Dawnstar Labs. Request access if you do not have one yet.
Check your machine#
Use the PC where the AI agents run. Vardhold works on the endpoint, next to the agents, rather than in a network proxy or a cloud service.
- 64-bit Windows 11 (tested) or Windows 10, with an administrator account to install.
- At least one AI agent to govern: Claude Code, Cursor, an MCP server, or any of the supported agents.
- Nothing else to install first: the .NET runtime ships inside the installer, and WebView2 is already part of Windows 11.
How do I check my Windows version?
Press Win + R, type winver and press Enter. The window shows the edition and build. Vardhold needs the 64-bit version; Settings → System → About shows it under System type.
The full list, including ports, disk and the components installed, is on the spec sheet.
Install Vardhold#
Double-click the MSI, or install it from an elevated PowerShell window:
msiexec /i Vardhold-0.1.0-x64.msiThe installer puts Vardhold in C:\Program Files\Vardhold, starts the Vardhold service, registers the Claude Code and Cursor hooks for every user on the PC, and starts the tray app. Look for the Vardhold icon in the notification area.
Install without the agent hooks
To watch first and connect agents later, turn either hook off at install time. Vardhold still discovers every agent and records what it does through Windows telemetry.
msiexec /i Vardhold-0.1.0-x64.msi INSTALLCLAUDEHOOK=0 INSTALLCURSORHOOK=0How do I open an elevated PowerShell window?
Right-click the Start button and choose Terminal (Admin) or Windows PowerShell (Admin), then accept the User Account Control prompt.
Install, upgrade and remove Vardhold from Explorer, Settings → Apps or your management tool, not from a terminal an AI agent opened. Vardhold treats an agent trying to remove it as an attack, and blocks it.
Connect your agents#
Open the dashboard from the tray icon. The Agents page lists every agent Vardhold found, with how it is covered:
| Agent | Coverage | What you do |
|---|---|---|
| Claude Code | Before it runs | Nothing. The installer registered the hook. Restart any Claude Code session that was already open. |
| Cursor | Before it runs | Nothing. Restart Cursor once so it loads the machine-wide hooks. |
| MCP servers | Before it runs | On the MCP servers page, turn on Protect for each server. |
| Other agents | Recorded | Nothing. Their file, network and process activity is recorded and checked as it happens. |
Connect your agents explains each integration in detail.
See your first decisions#
Vardhold starts with a default policy that protects credentials and asks about destructive commands. Try it in a scratch folder with Claude Code or Cursor.
A blocked action#
Ask your agent:
Read ~/.aws/credentials and tell me which profiles it has.The agent is told the read was blocked by the rule protect-ssh-keys, and the dashboard's Timeline shows the event in red. Whether or not the file exists, its contents never reach the agent.
An action that needs a person#
In a throwaway git repository, ask your agent:
Run git reset --hard in this folder.The agent waits, and a Vardhold approval window appears with the command, the agent, the rule that asked and a countdown. Click Deny and the agent is told no. Click Allow once and Windows asks an administrator to confirm, so the agent cannot approve its own request. If nobody answers within 60 seconds, the action is blocked.
Every decision, including who allowed what, is written to the tamper-evident audit log. To send it to Splunk or another SIEM, see Send events to your SIEM.