Dmitry Maslennikov

I’ve been building software products and startups for over 10 years. In that time I’ve launched products, managed teams, shipped features, and dealt with way more production incidents than I’d like to admit.

The problem that eventually turned into mttrly is simple: solo founders and small teams running their own servers spend a disproportionate amount of time on reactive ops work — SSHing in at 3am, reading logs they barely understand, applying fixes they’re not sure about. It’s not a skill gap. It’s a tooling gap. Enterprise tools like Datadog and PagerDuty are built for teams with dedicated SREs. Everything else is either too simple (UptimeRobot pings) or too risky (raw LLM with shell access).

In early 2026 I started building mttrly — a server management agent that operates through the dashboard, Telegram, and MCP-enabled IDEs, uses AI to diagnose incidents in plain English, and never executes a command without your explicit approval. It’s the tool I wish I had every time my phone buzzed at 2am.

mttrly is an AI-powered server management agent. You install a lightweight agent on your VPS, choose where approvals should reach you, and from that point you manage servers in plain English from the dashboard, Telegram, or an MCP-enabled IDE — without raw SSH as the normal control path, dashboard hopping, or 800-line log files to stare at.

The AI uses two models: Claude Haiku for fast real-time responses (sub-second classification and formatting) and Claude Sonnet for deep reasoning — reading logs, correlating events, diagnosing root causes. When something breaks, you ask “why is my site down?” and get a structured answer in 15 seconds, not a terminal prompt.

Safety is non-negotiable. The agent operates through scoped tools and a strict approval-gated, audited action layer — not raw unrestricted shell as the normal path. Any state-changing action (restart, rollback, config reload) requires your explicit approval before it runs. Dashboard approvals use passkeys; Telegram and MCP-enabled IDE approvals stay available under separate trust models. Everything is logged with the full AI reasoning chain so you can see exactly why each action was recommended.