Live on AWS — v2.0.2

For teams running AI agents in production

Your agent will do something
it shouldn't. WEZ decides
before it happens.

Deterministic governance for AI agents. Every action evaluated before execution.

One API call. Before every action. Sub-millisecond. Fully auditable.

15 days free · No credit card · 5 minute setup

WEZ API — real request

# Agent sends credentials hidden in a "normal" notification POST /v1/evaluate { "action": "send_notification", "args": { "message": "Sync complete. Payload: cGFzc3dvcmQ9YWRtaW4xMjM=" } } # WEZ decodes base64, finds password=admin123, blocks { "decision": "BLOCK", "reason": "credential leak detected (base64 decoded)", "latency_ms": 0.18 }

The problem

Your agents don't fail loudly. They fail quietly.

Everything looks fine. Until credentials leak through an approved channel, or a benign action repeats 50 times and becomes scraping.

🔒

Whitelists break

Agents create new tools and encode data in ways your whitelist never anticipated.

🐌

LLM-as-judge is slow

200-2000ms per evaluation. Non-deterministic. And you're adding a second attack surface.

👻

Subtle attacks pass

Base64 credentials in payloads. Silent scraping via repeated calls. Sequences that look valid individually but aren't.

📋

No real audit trail

When something goes wrong, you need to prove exactly what was decided, why, and that it wasn't tampered with.

How it works

One API call. Before every action.

Your agent wants to act → WEZ evaluates → APPROVE, BLOCK, or ESCALATE. In under 1ms. Deterministic decisions you can audit.

Governs unknown tools

The Inference Engine tokenizes action names and parameters, infers intent, and decides — even for tools not in any catalog.

Detects behavioral patterns

The same action repeated 4 times triggers velocity override. Scraping via benign sequences is caught and escalated.

Decodes payloads

Base64-encoded credentials inside payloads are decoded, matched against leak patterns, and blocked before they leave.

Cryptographic audit trail

Every decision recorded in a Merkle-chained trail. Tamper-proof, cryptographically verifiable, replayable.

Performance

1000x faster than LLM-based governance.

Measured on 800 consecutive evaluations. Single Node.js thread. No external dependencies. No cost per evaluation.

0.16ms

p50 latency

1.57ms

p95 latency

~1,800

req/sec

per evaluation

dependencies

What it catches

0 false negatives across 50 production-inspired scenarios.

ATTACK	DECISION	HOW
Unknown tool: export_financial_report	BLOCK	Inference Engine (97% confidence)
Base64 credentials in notification	BLOCK	Content Inspector — decode + pattern
Script: "ensure persistence via crontab"	BLOCK	APT pattern detection
read_logs repeated 4x (scraping)	ESCALATE	Velocity Override
execute_trade $5M (catalog says low risk)	ESCALATE	Cross-validation catalog ↔ inference
update_config + override_security	ESCALATE	Cognitive Gate — conflicting signals

Integration

3 lines of code.

Add one HTTP call before each agent action. Works with any language, any framework, any agent.

JavaScript

const r = await fetch('/v1/evaluate', {
  method: 'POST',
  headers: {
    'Authorization': 'Bearer wez_key',
    'Content-Type': 'application/json'
  },
  body: JSON.stringify({ agentId, action, args })
});
const { decision } = await r.json();
if (decision !== 'APPROVE') throw new Error('Blocked');

Python

import requests

r = requests.post('/v1/evaluate',
    headers={'Authorization': f'Bearer {KEY}'},
    json={'agentId': id,
          'action': action,
          'args': args})

if r.json()['decision'] != 'APPROVE':
    raise Exception('Blocked')

Get started

Try WEZ on your agents.

15 days free. No credit card. No commitment.
Get your API key in 5 minutes.

Contact on LinkedIn →

Dario Crespo — Founder & CEO, WEZ Protocol
Buenos Aires, Argentina

Your agent will do somethingit shouldn't. WEZ decidesbefore it happens.