OpenClaw
The free, self-hosted agent that works with any model.
OpenClaw is an open-source personal agent you run on your own computer or server, stewarded by the independent, non-profit OpenClaw Foundation. The software is MIT-licensed and free, with no paid tier. You pay only for the AI model you connect and, unless it runs at home, a server, typically $5 to $20/month for a small VPS.
How it works
OpenClaw is a single gateway process that connects your chat apps to an AI agent. You install it on your laptop, a small VPS or a server at home, pick a model and message it from Telegram, WhatsApp, Slack, Signal or any of its 29 supported channels. Memory, credentials and state live on your hardware.
The brain is your choice: GPT, Claude, Gemini or a local model that runs fully offline. You pay the model provider per token or through a subscription you already have. Agents resend long context at every step, so a loop can run up a bill quickly; watch your usage in the first weeks.
Safety is a matter of configuration. By default the gateway listens only on loopback, requires a token and sandboxes non-main sessions, and each tool can be set to allow, ask or deny. Rules live in openclaw.json and AGENTS.md, and the openclaw security audit command checks your setup. Stay current: anything before version 2026.4.22 has known critical flaws.
Strengths
- MIT-licensed, free, and no vendor account to create
- Works with any model, including local ones
- Talks through 29 chat channels, Telegram and WhatsApp among them
- Usable in countries no managed agent serves
- Memory and credentials stay on your hardware
Limits
- Every install, update and security setting is on you
- Open gateways and malicious skills have led to real breaches
- Model API bills can spike if the agent loops
- Onboarding is built for technical users
Technical users who want control over their data and model, and people in countries where no managed agent is offered.
Skip it if you do not want to maintain a server, apply updates or read through security settings.
Privacy
No company holds your memory or credentials, and with a local model nothing has to leave your machine. The main risk is your own setup: scans have found tens of thousands of gateways open to the internet, with many exposing API keys and tokens. Keep the gateway on loopback, reach it over SSH or Tailscale, and install as few skills as possible.
Plans that include it
| OpenClaw (self-hosted) Free software with no subscription, where you pay only for the model you connect and the machine that runs it. | Free |
Matchups
On the wire
- GPT-6.1 Sol cuts cached input to $0.10 per million tokens · September 29, 2026
The most flexible agent here and the only one you fully own, at the price of doing your own security work. Worth it if you are comfortable at a command line, or if nothing else is offered where you live.
Questions
Is OpenClaw really free?
Yes. The code is MIT-licensed and there is no paid edition. You only pay for the model behind it and the hardware it runs on.
How do I connect OpenClaw to Telegram?
Telegram is a built-in channel. Create a bot with BotFather, add its token to OpenClaw during onboarding, then message your bot and approve the pairing code for your own account.
Is OpenClaw safe to run?
It can be, if you stay on the latest version, keep the gateway bound to loopback, set shell commands to ask every time and install only skills you trust. Most real incidents involved exposed gateways or outdated versions.
Which AI models work with OpenClaw?
Any of the major ones, including GPT, Claude and Gemini, plus local models that run fully offline. Check separately that your chosen cloud provider serves your country.