Trust and operationsData handling

Data handling

AgentKeeper is designed to give security teams enough runtime evidence to govern AI agents without collecting more workstation data than the workflow requires.

AgentKeeper is designed to give security teams enough runtime evidence to govern AI agents without collecting more workstation data than the workflow requires.

What leaves the workstation

Depending on the installed source and the event emitted by that source, AgentKeeper can receive:

  • Source integration, such as claude_code, cowork, cursor, codex, mcp_gateway, or openai_compliance_api.
  • Workstation identity, such as hostname and local integration version.
  • Runtime event type, such as prompt, shell command, file read/write, web request, or MCP tool call.
  • Tool name, target path, command argv, domain, server name, or other event-specific metadata.
  • Policy verdict, reason, detector response, and matched policy context.
  • Redacted evidence used to explain a warning or block.

What is redacted

AgentKeeper should preserve the evidence needed to explain a policy decision while avoiding unnecessary raw payload capture. Examples:

  • File paths are useful for policy and investigation.
  • Secret values should be redacted.
  • Command structure is useful for detection.
  • Full command output is only retained when the product explicitly supports deep capture and the org has enabled it.
  • SSO and Directory Sync metadata should not be treated as runtime payload evidence.

Retention model

Retention depends on plan and feature settings. The practical admin rule is:

  1. Activity keeps the runtime audit trail.
  2. Deep Capture, when enabled, stores richer tool output for incident investigation.
  3. Prompt Vault decides whether raw prompts are visible, sealed behind audited reveal, or not retained.
  4. Policy and identity changes should remain auditable as administrative events.
  5. Deprovisioning a user should remove access, not erase historical security evidence by default.

Enterprise admin guidance

Before production rollout, document these choices:

  • Which AgentKeeper environment is production.
  • Whether Log All Tool Calls is enabled by default.
  • Whether Deep Capture is enabled.
  • Which groups receive stricter policy packs.
  • Who can administer SSO, Directory Sync, and policy.
  • How long the organization expects Activity and investigation evidence to remain available.
Was this page helpful?