AI Agent & MCP Security Architecture
How AgentGate enforces default-deny authorization, fail-closed MCP routing, request-bound approvals, secret redaction and tenant-scoped audit evidence.
Security boundary
AgentGate sits between an authenticated AI agent and the MCP server it wants to call. The gateway inspects the agent, server, tool, target environment, capabilities and arguments before forwarding the request. A direct connection to the upstream bypasses that boundary, so production deployment must make the governed route the required route.
Default deny
A call without a matching policy is blocked. Unknown agents and newly discovered tools do not gain implicit access.
Fail closed
If policy, persistence or approval checks cannot establish an allow decision, the upstream tool is not called.
Request-bound approval
One-time approval tokens are tied to the exact request, expire and reject replay or changed arguments.
Attributable audit
Authorization records connect tenant, agent, MCP server, tool, risk, policy and approval context.
Secret redaction
Pattern-based detection masks recognized secret-like values before audit storage. It is a safety layer, not a guarantee.
Tenant isolation
Management requests use tenant claims and RBAC; gateway credentials identify the tenant and agent for scoped access.
What AgentGate does not replace
Keep least-privilege IAM, network controls, database permissions, release approvals, endpoint security and SIEM monitoring. AgentGate adds the runtime authorization decision at the AI agent tool-call layer; it does not claim to eliminate prompt injection or certify an upstream MCP server.
Deployment and assurance
Managed HTTP MCP routing and customer-operated Docker / Helm deployment paths are available. Local Connector supports bounded local command-mode workflows. Long-lived stdio supervision, SSO, external WORM storage and independent security certification are not claimed as delivered capabilities.
For data retention, subprocessors, backup and recovery, responsible disclosure and operational commitments, read the AgentGate Trust Center. To verify the enforcement path yourself, run the protected MCP quickstart.
Create a live workspace. No production credentials required.
Plan a managed, private or self-hosted deployment.