OpenClaw, Installed and Actually Secure.

OpenClaw is an open-source, self-hosted personal AI assistant maintained by the OpenClaw Foundation, a non-profit, and released under the MIT license. You run a single gateway process on your own machine or server, and it becomes the bridge between your messaging apps and an always-available AI assistant. It was previously known as Moltbot and before that Clawdbot, so older tutorials you find may use those names.

The gateway is the core process, and it is the single source of truth for sessions, routing, and channel connections. One gateway instance serves every channel you have configured at the same time, so WhatsApp, Telegram, and Slack all talk to the same assistant rather than you running three separate bots. It also exposes a local dashboard and handles tool use, cron jobs, webhooks, and file, image, and audio handling.

More than 25 channels ship with it, including WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Google Chat, Microsoft Teams, Matrix, SMS through Twilio, IRC, LINE, Feishu, and Mattermost, with additional official plugins such as Nostr, Twitch, and Zalo installable on demand. There is also a plugin marketplace, so the list keeps growing. In practice most clients pick one or two channels and I connect those properly rather than switching on everything.

A modest cloud VPS is fine for most deployments. The main hard requirement is a current Node.js runtime - version 24.15 or newer is recommended, with Node 22.22.3+ and 25.9+ also supported. Storage grows with your session history, so give it a little headroom. If you would rather not manage a server at all, I can host it for you, or deploy onto a cloud account in your own name so you keep ownership of the infrastructure.

OpenClaw is not tied to one vendor. It works with Anthropic, OpenAI, Google Gemini, xAI Grok, OpenRouter, GitHub Copilot, and MiniMax, along with other compatible endpoints. That flexibility matters for cost as much as quality - routing routine work to a cheap fast model and reserving an expensive one for hard tasks is usually the difference between a sensible monthly bill and an alarming one.

Only if you set it up correctly, and this is the mistake I see most often. The dashboard binds to localhost by default for a very good reason - it is an interface to an assistant holding your credentials and message history. It should never be published straight to the open internet. The right approach is remote access over a private network layer such as Tailscale or an SSH tunnel, combined with sender allowlists so only approved contacts can message the agent, mention requirements in group chats, and a properly hardened host underneath. Every deployment I do includes that lockdown by default.

USER.md is the context file that tells your assistant who you are, what you are working on, and how you want it to behave, so every session does not start cold. It is the single highest-leverage file in the whole workspace, and it is also the one people fill with the wrong things. A good USER.md is specific, current, and short enough to stay useful; a bad one is a wall of detail that dilutes everything. I write yours as part of setup, and I have published a full guide with a real production example.

Your two costs are the server and the model usage. A small VPS is usually somewhere between free and roughly $30 a month depending on provider, and model usage is billed by whichever AI provider you connect. For a personal assistant handling a normal day of messages, model cost is typically modest; heavy automation costs more. I tune model routing during setup specifically to keep that number sane.

Yes. OpenClaw runs on macOS, Linux, and Windows, and there are companion apps plus voice support on macOS, iOS, and Android. Running it on your laptop is a perfectly good way to try it. The catch is that a laptop that sleeps, reboots, or travels is not an always-on assistant, which is why most people who get serious about it move the gateway to a server.

OpenClaw is a gateway first - its strength is connecting a very large number of messaging channels and plugins to an AI assistant. Hermes Agent, from Nous Research, is focused on the agent itself: self-improving skills, memory providers, subagents, and scheduled autonomy. Both are MIT licensed and self-hosted. I have run both in production and migrated a live agent between them, so I will give you a straight recommendation based on what you actually want it to do.

Yes. OpenClaw is MIT licensed, it runs on your hardware, and the configuration, USER.md, workspace files, and documentation are all yours. No license fee to me, no per-seat pricing, no lock-in. Ongoing maintenance is an optional monthly plan you can cancel whenever you want.