Getting started
Key concepts
The terms used throughout AgentID, from AI systems and verdicts to placeholders, restoration and failure modes.
AI system
An AI system is the unit you govern in AgentID: one browser rollout, one Claude deployment or one application that calls an LLM. Each system has its own guardrails, content policies, API keys, activity log and compliance records. Its System ID is a UUID you pass to the SDKs and APIs.
Integration target
The integration target says how a system is connected: Browser Extension, Claude Desktop, or SDK / API. It decides which setup the Integration tab shows and which settings apply. See Integration targets.
API key
An API key (sk_live_...) authenticates calls for exactly one AI system. AgentID stores only a hash and the key prefix; the full key is shown once when you create it. See API keys.
Guard check and verdict
A guard check is one call to the Guard API before a model call. Its verdict has three practical outcomes:
| Outcome | Response | What to do |
|---|---|---|
| Allowed | allowed: true, transformed_input equals the input | Call the model |
| Allowed with masking | allowed: true, transformed_input contains placeholders | Call the model with transformed_input |
| Blocked | allowed: false, reason set | Do not call the model; show or log the reason |
Placeholder
A placeholder is the typed token that replaces a sensitive value, in the form <TYPE_n>: <PERSON_NAME_1>, <EMAIL_1>, <PHONE_2>. The same value keeps the same placeholder within a request. AgentID uses typed placeholders only; it does not generate realistic fake values. See Placeholders.
Restoration
Restoration replaces placeholders in the AI's answer with the original values for the user. It happens on the user's device (browser extension, Claude integrations) or in your application (SDKs), never at the AI provider. It is off by default. See Restoration.
Active Guard and Shadow Mode
- Active Guard enforces blocking and masking before the model call.
- Shadow Mode never blocks or masks. Detections are logged as simulated outcomes (
simulated_decision:allowed,maskedorblocked) so you can calibrate guardrails on real traffic first.
Fail-open and fail-closed
When AgentID cannot complete a check (timeout, network error, dependency failure):
- Fail-open (default) lets the request continue. The Node.js SDK wrappers still apply their local masking of personal data and secrets.
- Fail-closed, enabled with Strict Security (Fail-Closed), blocks the request.
Shadow Mode always behaves as fail-open. See Failure modes.
Events and correlation
Each guard check and each logged answer is an event. A client_event_id ties the guard check and the answer of one model call together; reuse it as the event_id when you log the answer. Workflow steps of an agent share a workflowRunId instead. See Ingest API.
Activity log and Reveal
The activity log lists events per AI system. With Store Encrypted Payloads enabled, prompts, outputs and the originals behind placeholders are stored encrypted; admins can Reveal them, and every Reveal is recorded in the audit log.