AgentTrust ID

AgentTrust ID is a runtime authorization platform designed specifically for AI agents that execute frequent and potentially sensitive actions. The system addresses the need for per-action authorization, instant revocation, and secure delegation in environments where agents may delegate tasks to each other or interact across organizational boundaries. It operates by issuing opaque runtime tokens and evaluating agent actions routed through its SDK, MCP proxy, A2A adapter, or API session checks.

The platform classifies each action at runtime by matching action names against a set of 34 keywords distributed across four risk tiers: read, mutating, destructive, and admin. This classification determines the scrutiny level for each action, with unknown actions defaulting to a higher-risk tier. Session state is verified for each action, ensuring that the action is within the session's immutable scope and mode, and that agent ownership is validated. Every session is protected by HMAC-SHA256 integrity signatures. Behavioral analytics further enhance security, monitoring denial rates and write ratios to detect anomalies. If a session exhibits risky behavior, such as a high rate of denied actions or an unusual burst of write operations, the system automatically escalates the risk tier for subsequent actions.

AgentTrust ID employs a multi-tier Guardian pipeline for authorization decisions. Read actions are processed by the Fast Guard, while mutating actions are evaluated by both Fast and Spot Guards. Destructive and admin actions undergo the highest scrutiny, passing through Fast, Spot, and Deep Guards, with the Deep Guard leveraging AI reasoning via Claude for context-aware evaluation. The system is designed to fail closed for high-risk actions if any required Guardian tier is unreachable, ensuring no bypass is possible. Authorization decisions are returned with reasoning, confidence scores, and guard tier information, and all decisions are logged with SHA-256 chain continuity.

The platform supports cross-protocol authorization, allowing unified policy enforcement across MCP, A2A, and direct API integrations. Session-based privileges, delegation chain verification, and cross-organization federation are implemented to ensure secure and accountable agent operations. 0 license, with core services implemented as Go binaries and support for Redis and Postgres as backing stores. TLS and mTLS are configurable per deployment path, and an accountability API enables external verification of live session records.