Installation

AgentKeeper is installed per agent surface. Start with the agent the developer actually uses, confirm the workstation appears, then prove both PASS and BLOCK events in Activity before rolling out to more machines.

Use the Setup page from the same environment you plan to review:

• Production: https://www.agentkeeper.dev/setup
• Sandbox: https://sandbox.agentkeeper.dev/setup

The install command should contain a real AgentKeeper origin. If the command still shows a template placeholder instead of agentkeeper.dev or sandbox.agentkeeper.dev, do not run it.

On Windows, use browser/device approval for one workstation; use the Windows enterprise install path for fleets. Windows Quick Connect and the enterprise Windows path do not require Node or npm on endpoints.

On RHEL 8/9 or Rocky Linux 8/9 fleets, use the Linux RPM enterprise install path. The signed Universal RPM installs the runtime, pinned MCP Gateway, nine-surface catalog, systemd units, reconciliation, and offline repair assets. It does not install third-party IDEs or place an enrollment token in the image.

Windows Quick Connect

Run this from regular Windows PowerShell for one workstation:

irm https://www.agentkeeper.dev/windows-quick-connect.ps1 -OutFile $env:TEMP\agentkeeper-windows-quick-connect.ps1; powershell -NoProfile -ExecutionPolicy Bypass -File $env:TEMP\agentkeeper-windows-quick-connect.ps1

Quick Connect opens browser approval for the signed-in developer, then prompts for Windows elevation before installing the native AgentKeeper Gateway and Claude Code hooks. For many machines, use Windows enterprise install instead.

Claude Code Plugin

/plugin marketplace add rad-security/claude-code-plugin
/plugin install agentkeeper
/reload-plugins
/agentkeeper:connect

The reload step makes /agentkeeper:connect available in the current Claude Code session. The connect flow stores a device credential locally, registers the workstation, and installs hooks for prompt submission, pre-tool evaluation, post-tool audit, and session lifecycle events. Restart Claude Code after approval so those hooks load at startup.

Claude Chat in Claude Desktop

Claude Chat uses AgentKeeper MCP Gateway, not the Cowork plugin ZIP:

curl -fsSL https://www.agentkeeper.dev/install-gateway.sh | \
  AGENTKEEPER_API_URL=https://www.agentkeeper.dev \
  AGENTKEEPER_API_KEY=ak_live_REPLACE_ME \
  bash

On Windows, install the native Gateway binary first, then run:

& 'C:\Program Files\AgentKeeper\agentkeeper-mcp-gateway.exe' configure-ide --ide=claude-desktop

Restart Claude Desktop, start Gateway, and ask Chat to use a configured MCP tool.

Dashboard setup

1. Create an account at sign up.
2. Open Setup.
3. Choose the agent surface you want to connect.
4. Copy the command or hook configuration for that agent.
5. Complete the browser device flow or provide a managed API key if you are deploying through MDM.
6. Restart the agent after setup so startup-loaded hooks are active.
7. Run PASS and BLOCK validation.

Environment Variables

Supported agents

• Claude Code
• Claude Chat
• Claude Desktop / Cowork
• Cursor
• Codex
• Gemini CLI
• Windsurf
• GitHub Copilot Agent Mode
• Google Antigravity

Validation

A workstation is not considered fully connected until all of these are true:

1. A system or self-test event appears for the expected host.
2. A benign action appears as PASS when full tool-call logging is enabled.
3. A denied action appears as BLOCK.
4. The Activity source matches the integration being tested, such as claude_code, claude_chat, cowork, cursor, codex, gemini, copilot, windsurf, or openai_compliance_api.

See Configuration, Team Deployment, and Enterprise install for managed rollout patterns. Choose the Windows, macOS, or Linux package path first. Use Jamf, Iru/Kandji, and Intune pages as provider appendices for the platforms they manage.