Built for teams shipping AI agents under EU AI Act scrutiny. Sits between your agent and its tools, scores every call against your policy (allow, block, escalate), and writes a hash-chained audit trail an external auditor can verify without trusting your stack. The compliance report maps audit events to specific articles and flags gaps honestly instead of rubber-stamping them. Ships as a standalone MCP server or framework adapter. Chain heads anchor to RFC 3161 timestamp authorities for provable existence. The scoring blends rule-based gates with an optional XGBoost classifier trained on adversarial corpora, 140 microsecond hot path. Useful when procurement, regulators, or incident response demand cryptographic proof of what your agent actually did and why.
claude mcp add --transport stdio io.github.vaaraio-vaara-server uvx vaara-server