Skip to the content

Run AgentGarm on Windows

Install the Windows preview, start agents in its Job Object sandbox, use the tray, know what Windows enforces next to macOS and Linux, and remove it.

Built, not yet claimed preview: this guide describes a package that has not passed its release gate.

Install AgentGarm's Windows preview on a Windows 11 PC, start agents through its launcher so a Job Object, a restricted token and the Windows firewall hold them, follow denials from the notification-area icon, and know how Windows differs from macOS and Linux. Everything here is built and checked by builds and tests, but not yet run on a physical Windows 11 PC, and the build is not yet code-signed.

1. Install or upgrade the Windows package

Take the .msi for your PC's architecture and install it as the Windows install page shows. It installs per machine under C:\Program Files\AgentGarm and runs agentgarmd as the AgentGarm service under SYSTEM, which Windows restarts if it fails; agentgarmd then starts agentgarm-session in each signed-in session for you. State, the audit log and policies live in C:\ProgramData\AgentGarm, which only SYSTEM and administrators can read. To upgrade, install a newer release's package over it.

The build is an unsupported preview: it is not code-signed yet, so Smart App Control, where it is on, blocks it and SmartScreen warns before it runs, and every release carrying it is a pre-release. Verify the download first and use it to evaluate AgentGarm, not to rely on. For managed PCs, AgentGarm also packages the MSI for Intune, whose uninstall removes the hooks first, and as winget manifests.

agentgarm doctor

2. Install an agent's hooks and policy

From an elevated terminal, agentgarmd hooks install claude-code writes Claude Code's managed settings under C:\Program Files\ClaudeCode\managed-settings.d, and reports a conflict when a registry policy would shadow them; codex, cursor and the other agents take the same command. Put your policies in C:\ProgramData\AgentGarm\agents.d and start in audit mode, as on any other machine; agentgarmd reads them when it starts, so restart the AgentGarm service after an edit. Codex's hooks fail open on Windows as elsewhere, and on Windows they miss some of its patch edits and shell calls.

agentgarmd hooks install claude-code
agentgarmd hooks status claude-code

See also

Watch your agent for a week, then enforce

Start a policy in audit mode, see what enforcing it would have blocked in your own history, fix the rules that get in the way, then switch it to enforce.

3. Start agents through the launcher

agentgarm run places the agent in a Job Object that agentgarmd creates, with no way to break away, under a low-integrity, write-restricted token on a private desktop; on Windows 11 24H2 from build 26100.9278 it runs in an AppContainer of its own instead. agentgarmd holds the file guards, and Windows Filtering Platform filters keep the task's network to the local proxy. Every subagent and MCP server the task starts gets a nested Job Object inside the task's, whose memory, process and CPU limits bound the whole tree, and ending the task ends all of it. The launcher checks the sandbox with a canary at every launch and refuses to start the agent when it fails.

agentgarm run -- claude

4. Know what Windows enforces, next to macOS and Linux

In a launched session on Windows, files, the network, model calls through the proxy and resource limits all rate B, and approvals rate B with a security key enrolled. On macOS resource limits have no layer, and on Linux the network and model calls rate C. Command rules on the full argument list come only from the agent's hooks, rated C until each agent's live run is recorded, and MCP servers rate C. All of it is built, not yet claimed. Nothing rates A on Windows: its sentinel, an ETW trace of process starts, image loads and network connections tied to each task, watches and records but cannot block, so it cannot make launching through AgentGarm unavoidable. If something stops that trace, agentgarmd records it as tampering and starts it again.

Some things are only on other platforms for now: the computer-use airbag and the Endpoint Security sentinel are macOS only; Windows Hello stays switched off until its device tests are recorded, so approvals use a security key; agentgarm status does not say whether a Windows task's processes have ended; and a kill order can suspend or end a Windows task but not yet resume one.

See also

Approve risky actions on Windows

Enrol a security key on a Windows PC and approve each command your policy asks about through Windows' own WebAuthn; Windows Hello itself comes later.

5. Use the tray and check protection

agentgarm-agent tray puts the AgentGarm icon in the notification area, with Open dashboard and Quit, and shows a balloon for each denial, approval ask and alert, at most one every 6 seconds; to start it at sign-in, put a shortcut to it in your Startup folder. Without the tray, agentgarm notices prints the same notices in a terminal and agentgarm ui --open opens the dashboard. agentgarm status shows the services, each task's level and why, and agentgarm doctor checks the launcher's Job Object, restricted token, file guard and firewall filters, the agents' managed hook files and your presence key, and warns when run as SYSTEM.

agentgarm notices
agentgarm ui --open
agentgarm status

See also

Check that protection is on

See whether AgentGarm's services are up, what assurance level each agent session has and why, and what to do about a bypass alert.

6. Remove it

From an elevated terminal, run agentgarmd uninstall before removing the package: it takes out every agent hook and MCP configuration it installed, restoring what was there before, and removes the packaged helper registered for Windows Hello. If an install cannot be removed, it fails and names it. Then uninstall the package as the Windows install page shows. Policies, the audit log and the store stay in C:\ProgramData\AgentGarm.

agentgarmd uninstall