OpenClaw: a self-hosted AI assistant you message on Telegram/WhatsApp that does real work on your machine
TL;DR
OpenClaw is an agent runtime on your own machine (macOS/Windows/Linux) that connects a model (your API key or a local model) to files, scripts and the browser; you command it through chat apps like Telegram, WhatsApp, Slack, Discord. It has a community skill registry and Markdown-based memory. Run it sandboxed with least privilege.
Architecture in brief
| Component | Role |
|---|---|
| Gateway | Background process linking the model to local tools |
| Chat channels | Telegram, WhatsApp, Slack, Discord, iMessage… as the UI |
| Skills | Installable skill bundles, similar to Agent Skills |
| Memory | Context and preferences stored as editable Markdown |
What people use it for
- Texting “summarize unread email” or “remind me to review PRs at 4pm” from a phone.
- Scheduled scripts, report gathering, website monitoring.
- A personal assistant with long-term memory, not tied to one chat vendor.
Risks to know
- Shell access + inbound messages = big attack surface (prompt injection via messages, emails).
- Community skills can contain malicious code — read before installing.
- Run on a separate machine/VPS, sandboxed, without production credentials; restrict who can message the bot.
References: What is OpenClaw — DigitalOcean · OpenClaw on GitHub