Openclaw (topic)

OpenClaw is an open-source AI agent (software that acts on your behalf) that became the fastest-growing project of its kind, reaching 335,000 GitHub stars (a popularity measure for code repositories). In practice, it is the raw intelligence of a model plus every tool at its disposal, giving it full access to write code and take actions. Its creator, Peter Steinberger, was hired by Sam Altman in February to lead personal agents at OpenAI.

OpenClaw matters because its code, docs, and issues make the harness (the scaffolding around the agent) unusually visible, which is why engineers use it as a public case study for production AI systems. One feature, heartbeat, woke the agent every 30 minutes to check your stuff, so it started doing things that surprised you based on your goals. Vercel's CEO credited OpenClaw with showing how much a coding agent can do, informing his own product. OpenClaw 2.0 added multiplayer AI and strong Codex integration. It also shaped the field's direction: OpenAI endorsed third-party tools after Anthropic banned the project, a signal that AI tools are working together rather than against one another.

OpenClaw is a tool that runs an AI agent (software that acts on your behalf) you talk to in a chat. By default, OpenClaw ships in a single thread (one ongoing conversation), so if you wanted separate threads for different tasks, you had to set that up yourself. That was one of the things that made people quit. OpenClaw 2.0 fixed the thing that made users quit, and one of its breaking new features lets you ask your agent to visualize stuff. The chats can now do many things, including create their own widgets (small custom interfaces inside the chat) — so you can ask your agent to build dashboards or interactive things and show them inline.

Another big update in OpenClaw 2.0 is multiplayer AI (multiple people or agents in one session). You can literally have multiple instances of OpenClaw and connect them all together, though that feature is still baking over the next few weeks and the presenter planned a future video on how OpenClaw nodes work. If you're already running OpenClaw and it works, take your time before upgrading — maybe wait a week. A managed instance (a hosted setup someone else maintains) costs $5.99 a month and helps you skip the messy part.

OpenClaw is an AI agent tool (software that runs tasks for you). Its biggest documented pitfall is updates that break your setup. One reviewer noted that every time they updated OpenClaw, it would break, forcing them to open another agent such as Claude Code and ask it to fix OpenClaw — "the most annoying process of all." Another said they spent more time fixing OpenClaw than using OpenClaw, and that it "became wildly unreliable" before they switched elsewhere. A third reviewer observed a task hitting a 100% success rate outside OpenClaw but then "completely, completely broken" inside it, where the agent just stopped responding. The lesson: updates are risky, so test before trusting them.

OpenClaw 2.0 addressed this. It merged roughly 16,000 pull requests (proposed code changes) in one release — about half the project. Reactions split: some called it the best update yet, others said it broke their setup. One reviewer tested it on a brand new clean box (a fresh environment with no old configuration). That is the best practice: install fresh rather than upgrading in place. If your current OpenClaw already works, wait about a week before upgrading, so you avoid importing, migrating, or rolling back. A managed instance (a hosted setup someone else maintains) costs $5.99 a month and can skip the messy parts. OpenClaw 2.0 also improved visualizing results and chatting with widgets (small interactive panels) inside the chat.