AgentWall: Open-Source MCP-Proxy Runtime Safety Layer at 92.9% Enforcement Accuracy
Researchers have published AgentWall, an open-source MCP-proxy and OpenClaw plugin that intercepts every proposed agent action at runtime, evaluates it against a declarative policy, routes sensitive operations to a HITL gate, and produces an audit trail. Across 14 benchmark tests, AgentWall achieves 92.9% policy enforcement accuracy with sub-millisecond latency overhead. Compatible cross-host: Claude Desktop, Cursor, Windsurf, Claude Code, and OpenClaw.
Why It Matters
AgentWall is the first published, benchmark-validated instantiation of the "policy at the agent-action boundary" pattern — the enforcement layer that has been discussed conceptually for months but not yet concretely implemented and measured. The open-source release makes it directly evaluable against any MCP-based agent stack.