OPEN-SOURCE CLOUDFLARE WORKER · SELF-HOSTED

Verify Agent Intent
At the CDN Edge

Open-source Cloudflare Worker you deploy to your own account in five minutes. Patent-pending phonetic folding catches evasion attempts that regex misses. Pure TypeScript. Zero dependencies. Sub-millisecond. You own the keys, the logs, and the deployment.

Deploy in 5 MinutesRequest Source

Open-source. Runs on the Cloudflare Workers free tier (100K requests/month).

Why Regex Isn't Enough

Fraudsters misspell on purpose. Our phonetic folding normalizes Spanish dialect variants before pattern matching.

Seseo
c(e/i) → s
"hace" → "ase"
Yeismo
ll → y
"calle" → "caye"
B-V merger
v → b
"vamos" → "bamos"
H-deletion
h → ∅
"huevon" → "uebon"
D-weakening
VdV → VV
"pagado" → "pagao"
Z → S
z → s
"transferenzia" → "transferensia"
// All of these match the same pattern:
"dame tu contraseña"
"dame tu contrasena"
"dame tu contrazena"
"dame tu kontrasenia"
// All fold to: "dame tu contrasena" \u2192 BLOCK (social_engineering, weight: 40)

Three Commands to Deploy

# Get the source (email guard@merchantguard.ai for the tarball
# or wait for the public mirror, planned Q2)
tar -xzf merchantguard-cloudflare-llm-firewall.tar.gz
cd cloudflare-llm-firewall

# Authenticate with your own Cloudflare account
npx wrangler login

# Deploy to your account
npx wrangler deploy

# Your edge firewall is now live at:
# https://merchantguard-llm-firewall.<your-subdomain>.workers.dev
# All traffic, logs, and keys stay in your Cloudflare account.

Built for the Edge

🌐

CDN Edge Execution

Designed as a Cloudflare Worker. Deploy to your own Cloudflare account and run at every PoP Cloudflare operates (300+). Zero cold starts. Screens messages before they reach your origin.

⏱️

Sub-Millisecond

Deterministic pattern engine runs in linear time. No regex backtracking, no catastrophic blow-up on adversarial input.

🔊

Phonetic Folding

Spanish dialect normalization catches orthographic evasion that regex misses. "guevon", "huevon", "wuevon" all collapse to the same token.

🛡️

Prompt Injection Detection

"Ignore previous instructions", "forget your rules", "you are now". caught at the edge before your LLM ever sees them.

💰

Fake Payment Detection

"Ya te pague", "ya te hice la transferencia". pre-folded patterns catch misspelled and slang variants across 6 locales.

⚙️

Zero Dependencies

Pure TypeScript. No node_modules, no WASM, no external APIs. The entire verification layer ships as a single Cloudflare Worker bundle.

Edge Firewall vs. Sentinela API

Edge FirewallSentinela API
LayersPattern onlyFull cascade (4 layers)
Latency<2ms<10ms
Patterns25 (open source)161 x 6 locales
ML ModelsNoneXLM-RoBERTa + XGBoost
Trust AdaptationNoYes (4 presets)
Hallucination DetectionNoYes (15-dim vectors)
DeploymentYour CloudflareOur API
CostCF free tier$0.003/call
Best ForFirst verification layerDeep analysis

Use both together: Edge Firewall verifies 85%+ of traffic at the CDN. Sentinela handles the rest with full ML cascade. Learn more about Sentinela →

Pricing

Open Source

$0

100K requests/month on Cloudflare Workers free tier. Deploy in 5 minutes. 25 patterns.

RECOMMENDED

Enterprise

Custom

Full 161-pattern pack. Dedicated workers. On-prem option. SLA. Scales to $10K-$25K+/mo or $0.0005/req at volume. 14-day POC.

FAQ

How is this different from Sentinela?+
Can I use both together?+
What patterns are included?+
Can I add my own patterns?+
What about latency?+

Your LLM Deserves a Trust Layer.

Three commands. Zero dependencies. Every message screened at the edge.

Request Source (Free)Full Sentinela API

Patent Pending · U.S. App. Nos. 63/983,615; 63/983,621; 63/983,843