OpenClaw is free agent software you run yourself, whether on a spare laptop, a server at home or a machine you rent. It launched on 24 November 2025 and is looked after by the OpenClaw Foundation, a non-profit answerable to no lab. Its MIT licence comes with no premium edition and no hosted version, so the only bills are for the model you plug in and, if the agent does not live at home, a small VPS, usually $5 to $20 a month. It runs wherever Node.js does.
A gateway on your own desk
Technically, OpenClaw is one gateway program acting as a go-between for your messaging apps and an AI agent. Put it on a laptop, a machine under the stairs or a rented VPS, connect a model, and from then on you simply write to it in Telegram, WhatsApp, Signal, Slack or any other of its 29 supported channels. It can handle email, calendars, files and travel on your behalf. What it remembers, the credentials it holds and its working state all remain on your own equipment, and the brain can be GPT, Claude or Gemini, or a locally hosted model with no need for the internet at all.
The case for owning it
No vendor account, no subscription and no company holding your memories. You pick the brain and settle up with its provider directly, either by the token or via a plan you already pay for, and a local model can work entirely offline. It speaks through more chat channels than any rival and works in countries that no managed agent serves, which makes it the legitimate fallback for people in places such as Russia. Rules sit in plain files, openclaw.json and AGENTS.md, and every tool carries its own setting: allowed, ask first, or denied.
The price of freedom
Every installation, update and security choice falls to you. Early builds contained a bug that let an attacker run code with a single click; internet scans counted gateways sitting open in their tens of thousands, plenty of them spilling API keys and tokens; and booby-trapped skills have done the rounds. An outside audit has since been passed, yet every release before 2026.4.22 carries known critical flaws. Bills can surprise as well: because an agent sends its lengthy context again at each step, a runaway loop can inflate model charges fast. Onboarding presumes you are at ease in a terminal.
You are the security desk
Straight after installation the gateway accepts connections from the local machine alone, insists on a token and keeps secondary sessions in a sandbox. Leave those defaults alone. Connect from elsewhere through Tailscale or an SSH tunnel instead of exposing a port, make every shell command wait for your say-so, keep the list of skills short and trusted, and run openclaw security audit after each change. Pair it with a local model and nothing need leave your machine, a promise no managed agent can make, but that privacy lasts only as long as the configuration protecting it holds.
Suits
- Developers and confident tinkerers
- People who want to pick their own model
- Telegram, Signal and WhatsApp users
- Countries no managed agent serves
Look elsewhere for
- Anyone unwilling to maintain a server
- People who put off software updates
- Beginners hoping for polished onboarding
OpenClaw is the most flexible agent in the almanac and the only one you truly own, paid for in maintenance rather than money. It repays the effort if a command line holds no fears for you, or if no managed agent is sold in your country.