Skip to content
AgentIDDocs

Type at least two characters to search titles, headings and content.

↑ ↓ to navigate↵ to open
AgentID Docs

Getting started

How AgentID works

Every protected prompt goes through a guard check, an optional masked model call, an activity log entry and, if allowed, restoration on the user's side.

The request flow

User or application
   │  prompt, pasted text, files, tool results
   ▼
AgentID integration (browser extension, Claude plugin or companion, SDK, or your own HTTP call)
   │  sends the text to the Guard API
   ▼
Guard check ──► verdict
   │   • allowed            → continue with the text unchanged
   │   • allowed + masked   → continue with placeholders instead of sensitive values
   │   • blocked            → stop; the reason is shown or returned
   ▼
AI provider (receives only the protected text)
   │  answer, still containing placeholders
   ▼
AgentID integration
   │  logs the answer (Ingest API) and, if allowed, restores originals locally
   ▼
User sees the answer with the original values

1. The guard check

The integration sends the prompt to the Guard API together with optional context and up to four attachments. AgentID authenticates the API key, loads the AI system's guardrails and runs its detectors:

  • personal data and secrets (see PII detection),
  • prompt injection and jailbreak attempts,
  • toxicity,
  • database access and code execution patterns,
  • your content policies, including protected documents,
  • rate limits and token caps.

2. The verdict

The response says whether the call may continue (allowed), why not (reason), and the text to send (transformed_input). When masking is on, every detected value is replaced with a typed placeholder:

Original:   Send the contract to Jana Nováková, jana.novakova@example.com.
To the AI:  Send the contract to <PERSON_NAME_1>, <EMAIL_1>.

What happens with detected personal data depends on your settings: Mask replaces it, Block stops the whole request. See masking settings.

3. The model call

Only the protected text reaches the AI provider. The provider answers with the same placeholders, for example "Dear <PERSON_NAME_1>, ...".

4. Logging

The integration records the answer through the Ingest API. The activity log links the guard check and the answer into one row, so you can see what was checked, masked or blocked and what the provider received.

5. Restoration

If Restore masked values for the end user is enabled, the integration puts the original values back into the answer for the user:

  • the browser extension and Claude integrations restore them on the user's device,
  • the SDKs restore them in the response your application receives.

The AI provider and AgentID's logs keep the placeholders. See Restoration.

When AgentID cannot check

If AgentID is unreachable or a dependency fails, the system's failure mode decides: fail-open lets the request continue, fail-closed (Strict Security) blocks it. See Failure modes.