Kaylee — The Complete Guide
Who watches your server at 3 AM? An AI agent that never sleeps
Kaylee is the autonomous AI agent that keeps my entire agent network alive — around the clock, every day, without a break. Under the hood she runs on OpenClaw, a framework that has recently taken the agent world by storm: an AI agent that can perform almost any action you ask of it on a Linux server (start services, inspect logs, fix configuration, even edit source code) entirely on its own. OpenClaw pairs with Gemini Flash (Google's free tier) for reasoning, and ships with full access to the critical pieces of a host: containers (via Docker), system services (systemd), and the file system. That power is also its risk — it fires a lot of parallel requests at the model, so unconstrained use can get expensive fast. The fix is to set boundaries up front: an allowlist of permitted actions, protected paths, and budget guards. In my setup Kaylee speaks over Telegram (the @kylie_elad_bot), watches ten services at once, and only wakes me when she genuinely doesn't know what to do. For you she can replace on-call rotation, tidy up log noise automatically, or act as a general maintenance agent for any server environment that needs to stay stable.
What this guide covers
What is Kaylee? Meet your DevOps agent
OpenClaw running on Docker, reasoning with Gemini, chatting over Telegram
Kaylee is an autonomous AI agent with real hands on the server: she can run commands, check service status, edit configuration files, restart containers, and tail logs — exactly like a human sysadmin would. The engine underneath is OpenClaw, a relatively new agent platform built on top of the Claude Agent SDK that lets you wire any LLM into a Linux tool surface and let it work on its own.
Installing on an existing VPS
Her access starts from a single Docker bundle
Installing Kaylee means dropping one container onto an existing Linux VPS — the whole process rides on Docker (a system for running software inside an isolated 'box'), so you don't install dependencies directly on the host. On my setup she runs on Ubuntu 22 inside a roughly 200 MB container, and the journey from a fresh server to the first Telegram ping is about ten minutes. For you the same recipe works on any standard €5–€20/month VPS.
The toolbox she can reach for
From docker ps to grep-ing logs, all invoked in one sentence
Kaylee's toolbox is a curated set of 50+ prebuilt operations she can execute on the server — exactly the kind of commands a human sysadmin reaches for daily. The idea is not that she invents new commands, but that she picks from a well-defined allowlist (the formal term for a list of permitted operations) and fires the right one at the right moment. In my setup every action she runs is logged automatically, so I can rewind and see exactly what she did and when.
Scheduled monitoring — heartbeat + health + cleanup
Three layers of background surveillance
Scheduled monitoring is a set of jobs Kaylee runs on a fixed cadence in the background — without anyone having to ask. On Linux this happens via cron (the veteran scheduling system). In my setup three such layers run in parallel, each responsible for a different kind of check. Together they turn her from 'someone who answers when asked' into an agent that proactively looks after the server.
Self-healing
Hermes-style — detect, diagnose, repair, verify
Self-healing is Kaylee's ability to spot a problem, diagnose it, attempt a fix, and then verify the fix actually worked — all without needing to wake anyone up. This is the same pattern that Hermes implements across the whole network: from the first detection to the final alert when the system truly gives up. In operations parlance, escalation means handing a problem to someone (or something) that knows more. In practice it means Kaylee only pages me after she has genuinely tried and proven she can't handle it alone.
Safety: what Kaylee is not allowed to do
Allowlist, guardrails, and an audit trail
Safety is why an autonomous DevOps agent with full Docker access is a double-edged sword: the exact capability that lets her fix anything by herself also lets her destroy everything, accidentally or maliciously, if the wrong person takes control of her. In my setup Kaylee sits behind a whole stack of dos-and-don'ts defined in an allowlist (list of permitted actions), and every action is written to an immutable audit trail (tamper-evident action log). For you this section is non-negotiable before you let her anywhere near a production server.
Advanced tips
What I learned after four months of daily use
This section is the pile of nuances that separates 'I have a DevOps bot' from 'Kaylee is a genuine teammate'. These are things I (Elad) learned after four months of daily use on a real production server — tricks that multiplied her reliability, cut my model costs, and gave me tighter control over how she operates. For you, each one will save weeks of trial and error.
