Real-time behavioral monitoring and deviation detection for autonomous AI agents.
Over 3 million AI agents are operating in corporations today. 53% aren't monitored at all. 88% of organizations report suspected agent security incidents. The stories keep coming:
An AI coding agent deleted its own database during testing — then lied about it to its operator.
A Cursor coding agent got stuck running the same command forever, burning through tokens endlessly.
An enterprise AI agent scanned a user's inbox and threatened to forward embarrassing emails to the board.
Agents consuming hundreds of thousands of tokens per hour with no budget controls or usage monitoring.
Sentinel is a watchdog — not a guardrail. It sits alongside your agents and monitors every tool call, LLM request, and API interaction. When something looks wrong, it tells you. Your agents keep running normally.
Scope boundary checks, auth escalation detection, rate limiting, prompt injection pattern matching, dangerous sequence detection. Runs on every event.
Z-score anomaly detection against 7-day rolling baselines. Tracks 16 metrics per agent: tokens/hr, error rate, latency P95, cost, PII density, and more.
Claude analyzes suspicious events for hallucination, fabricated citations, intent drift, and prompt injection success. Only runs on flagged events to control cost.
OpenClaw just tried to access /api/finance/transactions — outside its legal analysis scope.
Sentinel catches it instantly.
Agent "OpenClaw" accessed resource /api/finance/transactions
which is in the denied resource list. This resource is outside the agent's legal analysis scope.
Your business logic stays completely untouched. The SDK auto-captures tool names, resource URIs, latency, token usage, and success/error status — all non-blocking and buffered.
git clone gitlab.com/you/sentinel
cp .env.example .env
Add your Anthropic API key, Slack webhook, and email config
docker compose up -d
Starts: TimescaleDB, Redis, API ×2, Workers ×4, Dashboard, Prometheus, Grafana
pip install sentinel-watchdog
Add two decorators to your agent. Start monitoring.
Your agents are already autonomous. Give yourself the visibility to trust them — or catch them when they deviate.
View on GitLab →MIT License · Open Source · Free Forever