LeSysBot is a small assistant that runs on one machine — usually a machine you own and occasionally need to reach when you are not sitting in front of it. You talk to it over a chat app, it talks to a language model running locally, and the model can call tools that do things on that machine: report temperatures, measure the link speed, check what is filling the disk, restart it, or power it off.
You: how hot is it running right now?
Bot: CPU 41–47 °C across 16 cores, GPU 38 °C. Nothing to worry about.
You: how much space is left on /?
Bot: 45 GB free out of 200 GB — 78% used.The whole system is three layers that know very little about each other:
| Layer | What it does | Swap it for |
|---|---|---|
| Messaging | Receives your text, sends replies | Terminal, Telegram, Discord, or an adapter you write |
| Agent | Runs the model, decides which tools to call | Any OpenAI-compatible endpoint |
| Tools | Does the actual work on the machine | Any folder with a README.md and a tool.py |
Because the layers are independent, the interesting question — "can it do X?" — almost always resolves to "write a tool". A tool is a folder. Drop it in, and it is live without a restart.
The decisions worth knowing up front#
It runs against a local model by default. The stock configuration points at Ollama on localhost:11434. Your messages, and the output of every tool call, stay on your hardware. You can point it at OpenAI instead by changing one URL — but then the machine's temperatures and disk usage travel to someone else's server, and that should be a decision you make deliberately rather than a default you inherit.
An install leaves three things running. A background service (systemd, launchd, or Task Scheduler — whichever your OS uses), the control panel it serves at http://127.0.0.1:8700, and a Grafana dashboard of the machine's CPU, memory, disk, network, and GPU. All of them bind to localhost only. Chat is the fourth thing, and the only one that reaches off the machine — which is why the allow-list below matters. Running lesysbot with no arguments prints the health of all four and exits without starting anything.
Destructive tools ask first. Anything that powers the machine down declares a confirmation string, and LeSysBot will not run it until you answer. The power tools go further and schedule the shutdown a minute out, so the reply reaches you before the network drops and you still have a window to cancel it.
Tools are ordinary Python that you are trusting. Installing a tool package from GitHub means running that author's code with your user's permissions on your machine. There is no sandbox. The trust model is exactly the same as pip install — read what you install, and prefer refs you have pinned.
Nothing needs root. No tool shells out through sudo or ships a sudoers rule, and the installer never asks for a password. Where a reading genuinely cannot be had unprivileged, the tool says so rather than elevating.
The allow-list is the security boundary. A Telegram bot with no allowed_user_ids will answer anyone who finds it, and it can power off your machine. The install wizard refuses to accept a blank list for this reason. See Security.
What it is not#
It is not a fleet manager — one install talks to one machine. It is not a general-purpose coding agent; the tools are deliberately small and single-purpose. And it is not a hosted service: there is no account, no server of ours in the path, and nothing to sign up for.
Where to go next#
If you want it running, start with Getting started — the wizard handles the model, the chat app, and the background service in one pass. If you would rather understand the machinery first, read How it works. If you already have it running and want to extend it, go straight to Write a tool.