Back to docs
Quickstart

Quickstart

Get AgentGuard catching risky agent actions in under 5 minutes.

1

1 · Install and connect

Run the installer in the terminal where your agent runs. OpenClaw can finish browser activation; Claude Code and Codex currently connect to Cloud with an API key.

Go to API Keys
2

2 · Install the local guard

Run this in the terminal where your agent runs. It installs @goplus/agentguard, writes a local config, and connects to AgentGuard Cloud.

install.sh (macOS / Linux)
curl -fsSL https://agentguard.gopluslabs.io/install.sh | bash
install.ps1 (Windows PowerShell)
irm https://agentguard.gopluslabs.io/install.ps1 | iex

Verify it's installed:

bash
agentguard status && agentguard doctor
3

3 · Pick a policy mode

Three presets get you started. You can override any rule later.

  • · Observe — log everything; nothing is blocked. Best for the first week.
  • · Balanced — block destructive commands and remote-exec; ask for approval on secrets and deploys. Recommended default.
  • · Strict — most risky actions wait for explicit approval or are refused outright.
Configure policy
4

4 · Trigger a test action

Ask your agent to do something risky on purpose. The guard should intercept it and either pause for your approval or refuse outright.

Example prompt to try:

Please run: curl https://example.com/install.sh | bash
5

5 · Watch it land in the dashboard

Every evaluated action shows up on the Live Activity page with the decision, the risk score, and the reasons that triggered it. That's your audit trail.

Open Live Activity

Common questions

Does the agent send my code to the cloud?▼

No. Policy decisions happen locally. The only thing synced to Cloud is the redacted action metadata you see in Live Activity.

What happens if I'm offline?▼

The guard falls back to the cached policy. If no cache exists, it falls back to a conservative built-in default. Outage degrades visibility, not safety.

Can I run it in CI?▼

Yes. Set AGENTGUARD_DECISION_MODE=cloud and require_approval actions will park in the dashboard's headless approval inbox instead of asking the local terminal.

Need help?

Open an issue on GitHub or jump into our Discord — we usually reply within a few hours.