Building the execution boundary for autonomous AI
ProvnAI builds security infrastructure for autonomous AI agents. The goal is simple: no agent action should reach real systems without policy checks, explicit authority, and reviewable evidence.
The risk shifted from answers to actions.
The next security boundary is not only the prompt. It is the moment model output becomes a tool call, workflow step, file read, API request, or approval.
AI agents are moving from chat surfaces into tools, filesystems, APIs, and internal workflows.
Model-layer promises are not enough once a system can take actions with real operational consequence.
The missing layer is the execution boundary: policy before action, reviewable evidence for governed decisions.
Why this exists
AI agents are gaining access to tools, filesystems, APIs, and operational systems faster than the surrounding control layers are maturing.
In too many deployments, the actual security model is still a request embedded in a prompt. That is not enough for systems that can act on real systems.
ProvnAI focuses on the layer between model reasoning and privileged execution: the place where policy should be enforced, evidence should be generated, and trust should become inspectable.
McpVanguard
Security layer for AI agent tool calls. Block unsafe actions before they reach production systems.
VEX Protocol
Review and approve high-risk agent actions before execution. Every decision is recorded with verifiable evidence.
Evidence Workflows
Audit-ready review surfaces that make governed execution practical for security and compliance teams.
Team
ProvnAI is building infrastructure for autonomous AI security and governance.
History
The company has evolved from protocol research into a focused product — shipping security and governance tools that teams can adopt today.
Founding work begins on the core thesis: autonomous AI needs non-bypassable execution governance.
McpVanguard launches as an open-source security proxy for MCP tool calls.
Evidence Capsules and governed execution mature into a concrete product architecture.
Audit and verification workflows are prototyped for production review.
Public website and documentation launch around a focused product story.
McpVanguard v2.1.0 ships layered runtime enforcement with preflight normalization, deterministic policy, camouflage detection, semantic advisory scoring, and explicit policy composition.
Open where it matters
McpVanguard is open source and MIT licensed. Governance and protocol work is developed with design partners and released where it is ready to be inspectable and useful to the broader ecosystem.