provnai
The trust layer theagentic erais missing.
Open-source infrastructure to secure, verify, govern, and audit
the next generation of autonomous AI agents.
Solving the
Black BoxProblem
Autonomous agents are inherently opaque. ProvnAI replaces blind trust with cryptographically verifiable traces of every decision, action, and state transition.
Proof of Execution.
Provnai transforms ephemeral agent logs into permanent cryptographic evidence. Every decision becomes a verifiable artifact.
Traditional Log
- Plaintext trace
- Easily modified
- No cryptographic link
Verifiable Receipt
- Merkleized proof
- Hardware signed
- RFC 6962 compliant
Evidence Portability
Portable receipts allow agents to carry their own proof, eliminating the need for blind trust between untrusted parties.
The Safety Stack
From hardware-level attestation to active tool-layer proxies. We consolidate fragmented agent security into a unified defense architecture.
McpVanguard
A security proxy for AI agents that use MCP (Model Context Protocol). It interposes between the agent and the host system, inspects every tool call, and blocks attacks before they reach your underlying servers.
Rules Engine
50+ YAML signatures — block path traversal, reverse shells, prompt injection, and SSRF attacks instantly.
Semantic Scorer
LLM-based intent scoring via OpenAI, DeepSeek, Groq, or Ollama to detect zero-day evasion attempts.
Behavioural
Shannon entropy and sliding-window anomaly detection. Stateful monitoring of conversational context.
Evidence Capsule[VEP]
Self-describing cryptographic container for agent logs. Every event (call, response, syscall) is Merkle-hashed and signed by a hardware-rooted TPM.
Adversarial Debate[RBD]
Cognitive verification protocol. Red/Blue agents debate the validity of a proposed action until a cryptographically verifiable consensus is reached.
Cognitive Routing[A2A]
Secure transport layer for agent-to-agent negotiation. Preserves temporal memory integrity and prevents context hijacking in multi-agent swarms.
VEX Protocol
The cryptographic substrate for the agentic era. VEX is a 17-crate Rust kernel that enforces a zero-trust security posture and mandatory auditability without sacrificing execution performance.
let capsule = VexCapsule::from_bytes(raw_data)?;
// Verify Merkle Root & TPM Signature
capsule.verify_integrity(tpm_pubkey)?;
match capsule.debate_consensus() {
Consensus::Allow => execute_action(capsule),
Consensus::Halt => panic!("Policy violation detected"),
}
Governed Execution
In most agentic systems, the component that proposes an action also authorises and executes it. Governed Execution mathematically separates these functions into independent primitives.
CHORA Gate Collaboration
The CHORA Gate holds continuation authority. Before any action executes, the agent requests a signed token. The VEX Authorization Enforcement Module (AEM) intercepts the syscall, verifies the token against a hardware TPM, and permits execution.
Evidence Capsules
The result is an Evidence Capsule — a cryptographically signed record of intent, authority, identity, and cryptographic witness — identically compatible between CHORA (Python) and VEX (Rust).
Authorization Enforcement Module (AEM) Handshake
Inference Proposes.
Governance Decides.
ProvnAI is co-authoring the .capsule Verifiable Agent Receipt specification alongside CHORA. We are defining the shared protocol for how autonomous agents prove their intent, authority, and identity across distributed ecosystems.
See It Run
We ran a 10x scale test pipeline using DeepSeek v3. The results verify VEX's concurrency model handles high-throughput agent swarms with minimal latency overhead.
Latency Comparison (Lower is Better)
VEP
Explorer.
Verify the cryptographic integrity of VEX Evidence Capsules locally. Zero server-side visibility. 100% cryptographic proof.
Live Logic Trail
Watch intent mapping as it happens.
Proof Validation
Every step is verified by the network.
Infrastructure Map
Quick links to all independent components of the Provnai ecosystem.
Where It Started
Before VEP. Before CHORA. Before Evidence Capsules.
VEXEvolve ran 29 autonomous agents for a full month — 480 articles researched, 158 published, 150 anchored to Solana. No human intervention.
That was VEX v0.1.4. A proof that verifiable autonomous agents work in the real world.
What we're building now is a different class entirely.
About the initiative
Provnai is an independent open research initiative. Everything published is open source under MIT or Apache 2.0 licenses.