Install and open
Load the controlled-release build, open a normal HTTP or HTTPS page, and click the Agent Guard toolbar icon.
Agent Guard docs
Authorize narrowly. Define intent. Review coverage.
Agent Guard is most useful when the user or administrator defines the authorized site, intended task, sensitive-data categories, destinations, budgets, and approval expectations before the agent acts.
Quick start
From installation to an explainable first review
Authorize one site
Grant access only to the current origin required for the workflow. Avoid blanket authorization.
Check Guard Health
Confirm which channels are guarded, observed-only, or unsupported before relying on approval or blocking.
Create an intent contract
Specify allowed sites, actions, categories, destinations, expiry, risk budgets, and user-presence requirements.
Choose a profile
Use Observe, Balanced, Strict, or a carefully reviewed Custom profile.
Review approvals and audit
Inspect the action summary, mismatch, prompt-risk association, coverage caveat, approval scope, decision, and result.
Core concepts
Use consistent definitions when reading the interface and evidence
Guard HealthCurrent coverage status for supported page actions and channels, including degraded or unsupported paths.
Intent ContractA bounded statement of what the user wants an agent to do, where, with which data categories, destinations, limits, and expiry.
Approval GateA pause before a supported high-risk commit, with minimized context and narrow approval choices.
Risk BudgetQuantitative limits for sends, external destinations, recipients, uploads, purchase amounts, irreversible actions, and account changes.
Data LineageA category-only graph linking source origin, sensitive category, supported action, and destination origin.
Injection TraceA time/source association between prompt-risk signals and later actions; it does not assert causation.
Session DiffComparison of destinations, categories, actions, prompt signals, and coverage between local sessions.
Panic LockAn emergency action that expires grants, cancels approvals, and blocks guardable high-risk actions within supported coverage.
Need help?
Include the product name, version, Chrome version, and minimal reproduction steps when contacting support.