Technical architecture and AI pipeline.
Central service in Node.js/TypeScript with WebSocket connections. Lightweight agent process on customer servers executes prepared playbooks and scoped local commands, then streams structured results back. PostgreSQL for audit logs and incident history. Hosted at app.mttrly.com.
The hosted configuration separates lighter intent and formatting work from heavier diagnosis and remediation planning. Current defaults use Claude Haiku and Claude Sonnet, while provider and model choices remain configuration rather than a fixed product promise.
The split reduces cost and latency while maintaining quality on reasoning-intensive tasks.
Every design decision assumes the AI can be wrong.
Scoped diagnostics and prepared playbooks are the normal path. One-off command execution is policy-gated and audited.
Risky changes normally create a pending action for a human to approve. A separately authorized Investigation session can allow a bounded set of command actions within server, time, and action-count limits.
The audit trail records operational requests, approval decisions, selected actions, and execution results. It does not promise access to a model's hidden chain of thought.
Unsupported requests and uncertain diagnoses remain visible to the operator instead of becoming an unrestricted shell action.
The current registry separates 36 read-only playbooks from 46 approval-required playbooks. Enforcement follows that runtime classification.
80+ built-in playbooks, currently 82 in the generated product reference:
Create an account with email, then install the required outbound agent on one server.