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.
TOOL execute_command
RISK CRITICAL
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.
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.
| Capability | Traditional WAF | API Gateway | mTrust 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.
Gateway
Every MCP request intercepted. Identity verified, trust score calculated, policy checked. <2ms.
Anomaly Engine
Batch analysis every 5 min. Frequency spikes, timing anomalies, trust cliffs, new agent bursts.
ML Intelligence
Parameter embeddings via Bedrock Nova. Sequence mining. Cross-agent correlation. Learns what "normal" looks like.
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
Every endpoint agent reports threat data. More devices → better signatures → fewer breaches for everyone.
20% of the internet flows through them. More sites → better bot detection → better protection for all.
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.
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.
AI agents managing infrastructure via MCP. Prevent unauthorized deploys, config changes, and privilege escalation.
Trading and transaction agents with behavioral oversight. Block anomalous patterns before they violate risk parameters.
Patient data access with per-agent trust verification. Audit every tool call. Detect data exfiltration attempts.
Ship MCP-powered products with security built in. Show customers their agents are monitored and policy-enforced.
Innovation Protected
mTrust Protocol is protected by pending patent applications covering our novel approach to decentralized, protocol-level AI trust management and behavioral verification.
A system and method for determining trust scores for autonomous artificial intelligence agents operating within a decentralized network protocol...
- URI-based interception methods
- Real-time behavioral scoring algorithms
- Context-aware policy enforcement engines