Vardhold · endpoint governance for AI agents

AI agents work with your developers' keys. We decide which doors they open.

Claude Code, Cursor, Copilot and MCP servers run commands and edit files with full user rights. Vardhold, from Dawnstar Labs, sees every agent on a Windows machine, stops the dangerous actions, asks a person about the risky ones, and keeps proof in your SIEM.

Get early access Run a 60-day pilot

Windows 11 · runs on the endpoint · no cloud account needed

The gap

Endpoint security sees processes. It does not see which agent did what, or why.

Agents act as the user

An AI coding agent can read SSH keys, browser profiles and .env files, push to production and change startup items, because the developer can.

Nobody approved it

Agent tools ask in their own windows, which settings or the agent can bypass. Auto-approve modes are one click away.

Nothing to show an auditor

When something goes wrong, there is no record tying an action to an agent, a rule and a decision.

What it does

Stop, ask, prove

One policy for every agent on the machine, enforced by a local Windows service.

block

Stop what should never happen

  • Reading SSH keys, cloud credentials, browser profiles
  • Turning off Defender or the firewall
  • Tampering with the controls themselves
ask

Ask a person about the risky

  • Force-push, hard reset, recursive delete
  • Persistence and startup changes
  • MCP tools that run shell commands

The person sees the exact command. Allowing needs Windows UAC.

record

Prove what happened

  • Tamper-evident, keyed audit log
  • Streams to Splunk or any SIEM, in OCSF
  • Kill switch for every agent at once

Coverage

Prevented where the agent allows it, detected everywhere else

AgentToday
Claude CodePrevented checked before each action
Any MCP serverPrevented every tool call and its arguments
CursorIn testing shell, MCP, file reads and edits checked before they run
VS Code Copilot, Copilot CLI, Codex CLI, Gemini CLI, Windsurf, local modelsDetected file and network activity; firewall and process control

Built for a security review

Trust that holds up

Approval agents cannot forgeAllowing an action goes through UAC on the secure desktop.
Fails closedIf a decision cannot be made, the action is blocked.
Local onlyYour data stays on your machines and in your SIEM. Nothing is sent to us.
Government-ready paperworkSBOM with every build, FIPS 140 statement, STIG-style hardening guide.

Pilot

See what your agents are doing within a week

  1. Week 1: install and discoverSilent MSI on 10 to 50 developer machines, in monitor mode: an inventory of every agent, MCP server and risky setting.
  2. Weeks 2 to 3: tune and enforceWe tune the policy with your team, connect your SIEM, then switch to enforce.
  3. Weeks 4 to 8: runBlocks and approvals in daily use, with a weekly check-in.
  4. End: reportAn AI agent risk report for security leadership, measured against criteria agreed at the start.

What the pilot includes

  • Vardhold on your pilot machines for 60 days
  • Hands-on setup and policy tuning with the founder
  • SIEM integration: Splunk, syslog over TLS or files
  • The agent risk report at the end
  • Pilot fee credited toward a first-year licence

Pre-release software. Windows 11 (64-bit). Pricing on request: [email protected]

Early access

Be first in line for Vardhold

Tell us a little about your team. We'll reach out about pilots and early access. No newsletter, no selling your details.

Or email [email protected]

We use your details only to contact you about Dawnstar Labs. See our privacy notice.