OpenClaw

OpenClaw is a free and open-source autonomous AI agent that runs on the user's own computer and communicates mainly through the messaging apps the user already has — Discord, Telegram, WhatsApp, Slack, Microsoft Teams, iMessage and more than 20 other channels — rather than through a dedicated web chat interface.^[1] It executes multi-step tasks such as browsing the web, summarising documents, managing calendars, drafting and sending messages, and scheduling automated routines, while keeping a persistent memory of past interactions.

History

OpenClaw was created by Austrian programmer Peter Steinberger and first published in November 2025 under the name Warelay. The software grew out of "Clawd", a personal assistant Steinberger had built for himself, whose name referenced Anthropic's Claude chatbot. Within three months the project was renamed twice: to Clawdbot in early January 2026, then to Moltbot on 27 January 2026 following trademark complaints from Anthropic, and finally to OpenClaw three days later because Steinberger felt the previous name "never quite rolled off the tongue".^[2]

The project went viral in early 2026 alongside Moltbook, a social network designed to be populated by AI agents, reaching roughly 247,000 GitHub stars and 47,700 forks by March 2026.^[2] On 14 February 2026 Steinberger announced he was joining OpenAI, and stewardship of the project passed to the OpenClaw Foundation, an independent US 501(c)(3) non-profit that employs the core team, signs every release and keeps the entire product MIT licensed, with no enterprise edition and no paid tier.^[3] Native iOS and Android applications, with chat, voice "Talk mode" and remote approval of agent actions, followed in June 2026, and Microsoft highlighted OpenClaw's native Windows runtime at its Build 2026 conference — the same week it launched Scout, a personal assistant openly inspired by OpenClaw.^[4]

How It Works

OpenClaw installs a local gateway on the user's machine that connects a large language model of the user's choice — the project is model-agnostic and is frequently paired with open-weight models — to messaging channels, a scheduler, file system access and tool integrations called skills. Because the agent runs on personal hardware, it can act on local files and calendars directly instead of copying them to a vendor's cloud. Its persistent memory stores facts and preferences across sessions so later requests build on earlier ones.^[5]

This architecture also creates the agent's principal risk: an agent that reads untrusted content, browses arbitrary web pages and can send messages is exposed to prompt injection. Steinberger has publicly acknowledged that the project "requires careful configuration to be secure", and IBM researchers described it as demonstrating that powerful agentic systems are no longer limited to large enterprises when given full system access.^[6] Security hardening — sandboxing, action approvals and least-privilege permissions — has therefore become a central community focus.

>See Also

• AI Agents
• Computer-Use Agents
• Model Context Protocol
• AI Memory
• Prompt Injection
• Open Weight Models