The First AI Firewall for MCP

Stop AI Agents
Before They Attack

mTrust is a managed AI firewall that sits between your MCP servers and the agents that call them. Real-time trust scoring, behavioral anomaly detection, and policy enforcement — deployed in minutes.

mtrust-gateway — live
Gateway Active — Proxy Mode
-- Incoming MCP Request --
AGENT claude-3.5-sonnet
TOOL execute_command
RISK CRITICAL
⚠ Trust score 42 — threshold 95
✗ DENIED — anomaly: parameter deviation detected
Agent blocked. Alert sent to dashboard.
Evaluations
24.3K
Last 24h
Anomalies
7
3 blocked
AI Firewall Active

3 MCP servers protected · 12 agents tracked · <2ms latency

AI Agents Are Calling Your Tools.
Nothing Is Watching.

A new protocol called MCP (Model Context Protocol)is becoming the standard way AI agents interact with software. When ChatGPT reads your files, when Claude executes a database query, when an AI assistant deploys your code — they're using MCP to call tools on servers you control.

The problem:MCP has no security layer. When an agent connects to your MCP server, there's no identity check, no trust score, no behavioral analysis, no audit trail. The server just executes whatever the agent asks. If that agent gets compromised — through prompt injection, a supply chain attack, or a malicious model — it has full access to every tool on your server.

This isn't theoretical. In 2025, compromised AI agents autonomously exfiltrated data from Fortune 500 companies with 80-90% autonomy. As MCP adoption accelerates — Anthropic, OpenAI, Google, and hundreds of startups are building on it — every MCP server deployed is another unprotected entry point.

mTrust is the firewall for MCP.It sits between AI agents and your tools, intercepting every request. It identifies the agent, calculates a real-time trust score based on behavior, enforces your security policies, and learns what “normal” looks like — so it can catch attacks that rule-based systems miss.

Think of it as CrowdStrike for AI agents
No Security Layer Exists

MCP Has No Firewall

Networks have firewalls. Web apps have WAFs. APIs have gateways. But when an AI agent calls your MCP server, nothing checks who it is, what it's doing, or whether it should be allowed. Every MCP server deployed today is an open door.

It Already Happened

Compromised AI agents autonomously exfiltrated data, executed unauthorized commands, and pivoted through connected systems — with 80-90% autonomy and zero human oversight.

Prompt Injection → Full Compromise

Any agent with MCP access can be hijacked to read files, execute commands, and exfiltrate data through the tools it already has permission to use.

Your WAF Can't See It

Traditional firewalls and API gateways approved the traffic because it looked like valid API calls. They have no concept of agent identity, trust, or behavioral patterns.

MCP Adoption Is Accelerating

Anthropic, OpenAI, Google — every major AI lab is converging on MCP. Every server deployed is another unprotected entry point.

Agents Are Moving to Production

Code execution, database access, infrastructure management — AI agents are no longer demos. They're running real workloads with real consequences.

No One Is Checking

MCP has no built-in authentication, no trust scoring, no behavioral analysis, no audit trail. The protocol maintainers have shown no indication of adding one.

mTrust Fills the Gap

The first and only firewall purpose-built for the MCP protocol layer.

“Can't We Just Use a WAF?”

Traditional security tools were built for HTTP traffic from humans. They have no concept of AI agent identity, trust, or behavioral patterns.

CapabilityTraditional WAFAPI GatewaymTrust AI Firewall
Understands AI agent identity
Per-agent trust scores
Behavioral baselines
Per-agent statistical models
Tool-level policy enforcement
Risk-tiered per tool
Parameter semantic analysis
Embedding similarity (Bedrock Nova)
Sequence attack detection
Markov chain analysis
Coordinated attack detection
Cross-customer correlation
Adaptive severity scoring
Trained on operator feedback
Natural language explanations
Claude-powered
MCP protocol native
mtrust:// URI scheme

An AI Firewall That Gets Smarter

Not just a gateway — an adaptive security layer that learns what normal looks like for every agent, detects novel attacks, and responds autonomously. The more customers deploy it, the better it gets for everyone.

Intercept Every Request

The gateway sits between agents and your MCP server. Every tool call is intercepted, the agent is identified, and a trust score is calculated — in under 2ms.

Learn What Normal Looks Like

Per-agent behavioral baselines build automatically. Parameter embeddings via Bedrock Nova detect when tool arguments deviate from what this agent normally sends.

Detect Novel Attacks

Sequence mining catches unusual tool call patterns. Cross-agent correlation detects coordinated attacks across multiple customers. Claude explains what happened in plain English.

Respond Automatically

Configurable auto-block rules, auto-resolve for low-risk anomalies, webhook integration, and anomaly-triggered policy adjustments. The system acts while you sleep.

See Everything

Full dashboard: server health, agent trust scores, anomaly triage, audit logs, billing. Every decision logged. CSV/JSON export. 30-second auto-refresh.

Deploy in Minutes

Proxy mode: point agents at us. Sidecar mode: deploy alongside your server via Docker, ECS, or Kubernetes. Protected MCP server in under 10 minutes.

Three Layers of Defense

Every MCP request passes through the gateway. Suspicious patterns are caught by the anomaly engine. ML intelligence learns what normal looks like.

1
Intercept

Gateway

Every MCP request intercepted. Identity verified, trust score calculated, policy checked. <2ms.

Detect

Anomaly Engine

Batch analysis every 5 min. Frequency spikes, timing anomalies, trust cliffs, new agent bursts.

Learn

ML Intelligence

Parameter embeddings via Bedrock Nova. Sequence mining. Cross-agent correlation. Learns what "normal" looks like.

Respond

Allow / Deny / Escalate

Decision logged. Agent score updated. Anomalies surfaced in dashboard for triage.

Every Customer Makes Everyone Safer

mTrust isn't just a product — it's a sensor network. Every deployment adds behavioral data to a shared intelligence layer. Attack patterns discovered at one customer protect all others before they ever see the threat.

More Customers

Each deployment adds interaction data to the global training pipeline

Better Models

More data → better parameter centroids, sequence models, severity predictions

Fewer False Positives

Better models → more accurate detection → happier customers

Network Immunity

100 customers is a sensor network. 10,000 is an immune system.

The Same Playbook That Built $80B+ Companies

CrowdStrike
Endpoint telemetry

Every endpoint agent reports threat data. More devices → better signatures → fewer breaches for everyone.

Cloudflare
Web traffic patterns

20% of the internet flows through them. More sites → better bot detection → better protection for all.

mTrust
AI agent behavioral data

Every MCP interaction builds the model. A dataset no one else is collecting — at the protocol layer where it matters.

A competitor can copy the gateway rules. They cannot copy the global behavioral models trained on millions of real AI agent interactions.

See the Dashboard

Monitor every MCP server, agent, and anomaly from a single dashboard. Built with Next.js, powered by real-time data.

app.modeltrust.io/dashboard
Evaluations (24h)
24,312
↑ 12% vs yesterday
Allow / Deny / Escalate
94% / 5% / 1%
Active Agents
47
across 3 servers
Anomaly Alerts
7
3 high severity
prod-api-serverhealthy
8420 evals today
internal-toolshealthy
12100 evals today
data-pipelinehealthy
3792 evals today

7 Pages, 6 Server Tabs

Home, servers, server detail (overview, tools, policies, agents, audit, settings), audit log, billing, settings.

Add Server in 4 Steps

Enter origin URL → auto-discover tools → set risk levels → deploy. Protected MCP server in under 10 minutes.

Real-Time Anomaly Triage

Filter by severity, type, server, agent. Bulk dismiss or block. 30-second auto-refresh. CSV/JSON export.

Built on AWS, Designed for Scale

Production infrastructure with tier-based scaling, managed by us so you don't have to.

Platform

Trust Evaluation
<2ms
Anomaly Detection
5 min batch
Compute
ECS Fargate
Database
DynamoDB
ML Embeddings
Bedrock Nova
Auto-Scaling
CPU 70%, max 10

Security

Auth
AWS Cognito
API Keys
Secrets Manager
Encryption
AES-256 at rest
Audit Trail
Every request
Anomaly TTL
90 days
Alerts
SNS real-time

Who Uses mTrust

Any team deploying MCP servers needs to know which agents are calling them and what they're doing.

DevOps & Platform

AI agents managing infrastructure via MCP. Prevent unauthorized deploys, config changes, and privilege escalation.

Financial Services

Trading and transaction agents with behavioral oversight. Block anomalous patterns before they violate risk parameters.

Healthcare AI

Patient data access with per-agent trust verification. Audit every tool call. Detect data exfiltration attempts.

AI-Native Startups

Ship MCP-powered products with security built in. Show customers their agents are monitored and policy-enforced.

Patent Pending US 63/839,767

Innovation Protected

mTrust Protocol is protected by pending patent applications covering our novel approach to decentralized, protocol-level AI trust management and behavioral verification.

First protocol specifically designed for AI agent security
Addresses documented real-world threats (Anthropic Claude attack)
Validated by cybersecurity experts and enterprise customers
Application No.
US 63/839,767
Abstract

A system and method for determining trust scores for autonomous artificial intelligence agents operating within a decentralized network protocol...

Claims Coverage
  • URI-based interception methods
  • Real-time behavioral scoring algorithms
  • Context-aware policy enforcement engines