AAV open-sources a runtime security layer for AI agent actions with MCP approvals

CarlosMarreroAAV · reddit · 2026-09-22

Developer Carlos Marrero released AAV, a runtime security layer for AI agent actions, now open source under Apache-2.0 after pivoting from a commercial product. Core tenet: agent security boundaries shouldn't depend on proprietary clouds. Features include controlling executable actions, human approval for consequential operations (MCP approvals), credential isolation, and verifiable audit receipts. AAV Cloud remains optional for centralized management, but the local enforcement layer is independently usable. Early 0.1.0-alpha, seeking developers and security folks to stress-test it.

Original post →

More from coding & agent

coding & agent channel →