OpenClaw

OpenClaw is an open-source personal AI agent you host yourself. Instead of shipping another chat window, it runs as a gateway on your own machine — macOS, Linux or Windows — and exposes that agent through the messaging apps you already keep open: WhatsApp, Telegram, Discord, Slack, Signal, iMessage and roughly two dozen more channels, in DMs and group chats alike. Because the process lives on your hardware, it can read and write local files, run shell commands, drive a browser and hold memory across sessions rather than starting cold every time.

The project started as a weekend build by Peter Steinberger (founder of PSPDFKit) under the name Clawdbot in late 2025, was renamed OpenClaw in January 2026, and became one of the fastest-growing repositories in GitHub's history. It is now stewarded by the non-profit OpenClaw Foundation with a full-time maintainer team, and the code stays MIT-licensed on npm as the openclaw package.

OpenClaw 2.0 (v2026.8.1, released 30 August 2026). The largest release in the project's history — roughly seven weeks of work, 16,000+ merged pull requests from 933 contributors, 569 of them first-timers, which the maintainers put at about half of every PR ever merged into the project. Three changes matter most. Setup got much shorter: the installer now detects what is already on your machine — an existing Claude or ChatGPT subscription, API keys, locally served models — and moves the remaining configuration into conversation with the agent afterwards rather than a wall of prompts up front. The browser app is now the primary surface, not a secondary one: you land in a conversation with your agent and do setup, monitoring of running work and live activity from the same place. And shared cloud sessions make it multiplayer — live sessions can move between the gateway and runners or onto a paired device, several people can join a session with its context intact, and the Control UI disambiguates multi-account session labels and records who archived a collaborative session. Security work in the same release tightens gateway authentication (unauthenticated network gateway installs are now blocked), adds safe gateway token recovery during onboarding, clears stale remote passwords when switching to token auth, and introduces an explicit limited mode for mobile pairing alongside full access. Messaging, memory, skills, models, automations, the native apps and plugins were all touched too.

How it is set up. A single install script (curl -fsSL https://openclaw.ai/install.sh | bash) provisions Node and dependencies; from there you pair chat channels and grant capabilities one at a time. A companion iOS app pairs with your gateway by QR code so an iPhone can act as a secure node for chat, voice "Talk" mode, action approvals, and device context such as camera, location, photos, calendar and reminders.

Models and cost. OpenClaw is bring-your-own-model. It works with Anthropic's Claude, OpenAI's GPT family, and locally served open-weight models, and it can ride an existing Claude or ChatGPT subscription instead of metered API keys. The software itself is free; what you pay is whatever inference you point it at.

Worth knowing before you install. An assistant with shell access, file access and your messaging accounts is a large amount of trust in one process. OpenClaw ships an approval flow for sensitive actions, but self-hosting means you own the security posture — sandboxing, credential scope and channel access are yours to configure, not the vendor's to guarantee.