Prompt-injection guard
Detects override attempts, jailbreak personas, hidden instructions, encoded payloads, role hijacking, delimiter collapse, and output leakage before they reach the agent pipeline.
Mona.Expert NGO · Köln, NRW, Germany
AI agents should ask before they touch reality.
mona.expert is a security wrapper and control plane for autonomous AI agents. It sits between an agent and tools, APIs, files, databases, websites, and outbound network calls — then applies policy, detection, approvals, and audit logging before anything happens.
mTLSgateway-to-wrapper authentication
Audithash-linked, tamper-evident logs
Policydeny-by-default action control
Localself-hostable open-source core
What mona.expert does
The wrapper gives agents a controlled perimeter. The agent can request an action, but the gateway decides whether the action is allowed, held for review, denied, or quarantined.
1 · ORG HALT
2 · CAPABILITY
3 · RATE LIMIT
4 · POLICY
5 · DETECTION
6 · AUDIT
Tool permission model
Each agent gets an explicit toolset. Unknown tools are blocked. Risky tools can require human confirmation, dry-runs, or approval from the dashboard.
mTLS backend split
The public website gateway on port 4188 talks to the wrapper backend on 4189 with client certificates. Random local processes cannot talk to the backend.
Live device overview
The dashboard displays OS, hostname, CPU, memory, Node runtime, sessions, actions, risk scores, and event history in a compact low-overhead interface.
Audit trail
Actions are written to append-only JSONL/audit records with hash chaining so later edits become visible. The web dashboard can show live feed, counters, and evidence.
Website connection
The wrapper can connect to the hosted website dashboard for account management, multi-agent overview, setup guidance, and team-facing status pages.
Live security check
This local demo calls /api/security-check on your running wrapper. Try a normal instruction or a prompt-injection attempt.
Ready
Click “Run check” to test the local wrapper API.
Website auto-connect status
This panel reads the real local wrapper state. It does not fake cloud connection. If the wrapper is connected to a mona.expert account, you will see the actual wrapper id and connection time.
Local wrapper
Checking…
Reading /healthz and /v1/connect/status.
mona.expert website
Checking…
No cloud state shown until the wrapper reports it.
Link it
npx mona-expert connect <api-key-from-dashboard>
Local setup
Use the local wrapper when you want the dashboard and API on your own machine. Use the hosted website when you want team login, remote overview, or deployment status.
quick start
cd ~/mona.expert
npm install
node wrapper-server.js
node website-server.js
open http://127.0.0.1:4188/index.html
Interfaces
| Path | Purpose | Status |
|---|---|---|
/index.html | Product homepage and live security check. | Local + hosted |
/dashboard.html | Hosted account dashboard and wrapper status overview. | Website UI |
/dashboard/ | Compact local control-room dashboard with gate tape, sessions, agents, and risk charts. | Local UI |
/api/security-check | Checks text for prompt-injection and unsafe control signals. | JSON API |
/v1/dashboard/summary | Local dashboard summary: system metrics, agents, sessions, actions, alerts. | JSON API |
/healthz | mTLS-backed health check through the website gateway. | JSON API |
Company
mona.expert is maintained by Mona.Expert NGO, based in Köln, NRW, Germany. The mission is practical AI safety infrastructure: simple enough for individual builders, strong enough for teams that need accountability.
Mona.Expert NGO
Location: Köln, North Rhine-Westphalia (NRW), Germany.
Focus: AI agent security, transparent auditability, self-hostable infrastructure, and safe automation workflows.
Operating principle
Agents should not directly hold secrets, execute high-risk tools, or silently bypass oversight. Every real action should be visible, attributable, and policy-governed.
FAQ
Is this only for OpenClaw?
No. OpenClaw can connect to it, but the wrapper is designed as a generic perimeter for AI agents and tool-calling systems.
Does the browser store API keys?
No. The design avoids JavaScript-readable credential storage. Website authentication should use httpOnly cookies or short-lived server-issued sessions.
Can it run offline?
The local wrapper and dashboard can run on your machine. Hosted account features require the website and database.