Troubleshooting & FAQ
FAQ
Short answers to the questions teams ask most often about AgentID.
Does the AI provider ever see the original values?
Not when masking is on and the call goes through AgentID: the provider receives placeholders such as <PERSON_NAME_1>. Calls that bypass AgentID, or checks that fail open without local masking, are not protected. See Failure modes.
Does AgentID see my prompts?
Yes, Guard must read a prompt to detect sensitive values. What it stores afterwards is described in What is logged; with Store Encrypted Payloads, stored content is encrypted and readable only through an audited Reveal.
Can AgentID replace names with realistic fake names instead of placeholders?
No. AgentID uses typed placeholders only, because they can always be restored exactly. See Placeholders.
Where are masked values restored?
On the user's device (browser extension, Claude plugin and companion) or in your application's response (SDKs), and only when Restore masked values for the end user is on. See Restoration.
Should I block or mask personal information?
Mask when users need to keep working with the AI on real cases; block when such data must not be used with AI at all. Blocking stops the whole request.
What is the difference between Shadow Mode and fail-open?
Shadow Mode is a deliberate setting: nothing is blocked or masked, outcomes are only simulated. Fail-open is what happens when a check cannot complete in Active Guard: the request continues.
Which AI web apps does the browser extension protect?
ChatGPT, Claude, Gemini and Microsoft Copilot. Other AI apps are discovered and can be warned or blocked, but their content is not checked. See Supported AI apps.
Does the browser extension work in Edge?
Yes. It is a Chromium extension for Chrome and Edge.
Do I need both the Claude plugin and the Claude Desktop companion?
For full coverage, yes. The plugin protects Cowork and Claude Code; Claude does not run plugins in ordinary chats, which the companion protects.
Which languages does detection support?
Validated formats (e-mail, phone, IBAN, card numbers, national IDs, API keys) work in any language. Detection of names, organizations and addresses in free text is tuned for English, Czech and Slovak.
Is there a Windows endpoint agent?
Windows endpoint discovery is not offered for new AI systems. Use the browser extension for AI web apps and the Claude integrations for the Claude app.
Which SDK versions does this documentation describe?
agentid-sdk 0.1.43 (Node.js), agentid-vercel-sdk 0.1.37, agentid-sdk 0.1.27 (Python), browser extension 0.1.12, Claude plugin 0.9.6, Claude Desktop companion 0.5.4 (Windows) and 0.2.1 (macOS).