Skip to main content

Agent Security

A gateway for AI agents. Every tool call passes through enforcement before it reaches the upstream server.

How it works​

Agents connect to a single endpoint. WardenAuth fans out tools/list across all your MCP servers, namespacing each tool (github__list_issues, slack__post_message). No collisions. Every tool call passes through five checks:

  1. Consent — Is this tool + tier in the agent's grant?
  2. DLP — Do the arguments contain secrets or prompt injections?
  3. Velocity — Is the agent within its rate budget?
  4. Resource constraints — Is the agent accessing allowed data?
  5. HITL — Does this tool require human approval?

Key capabilities​

Two-level HITL​

  • Self-approval — the agent's owner approves from their phone
  • Admin-approval — an administrator must approve, enforced by SoD

Approvals arrive via Web Push with deep links — approve from anywhere.

Drift detection​

If an upstream MCP server adds tools or changes descriptions, WardenAuth detects it and raises an alert. No silent permission creep.

Cryptographic receipts​

Every allow/deny produces a signed receipt. Verify it independently — see Permission Receipts.