What Is OpenClaw? The Open Source AI Agent Formerly Called Clawdbot and Moltbot

OpenClaw is a free, open source AI agent that runs on your own computer and takes instructions through the chat apps you already use, such as WhatsApp, Telegram, iMessage, Slack and Discord. You message it like a person; it can browse the web, run commands, handle email and calendars, and do things on a schedule without being asked. It doesn’t come with a model: you connect one, either a cloud provider or a local model.

Clawdbot, Moltbot, OpenClaw: the same project

If you’ve heard of Clawdbot or Moltbot, that was OpenClaw. Peter Steinberger first published it as Warelay on November 24, 2025. According to Wikipedia’s history of the project, it was briefly CLAWDIS in December, became Clawdbot on January 2, 2026, was renamed Moltbot on January 27 after trademark complaints from Anthropic, and became OpenClaw three days later because “Moltbot” didn’t roll off the tongue. Old guides and forum threads under any of those names describe the same software, though much has changed since.

Today the project is stewarded by the nonprofit OpenClaw Foundation and is MIT licensed. There is no paid tier, no hosted service and no company selling an upgrade.

What it does

The core idea is a single always-on agent that lives where you already talk. Per its integrations page, it covers 29 chat channels, works with 64 model and media providers, and has 142 official plugins plus a large community skills ecosystem. It keeps memory across conversations, runs scheduled jobs, and has a default “Proactive check” job that wakes it up to look for things worth telling you (heartbeat docs). You can run several agents and route different chats to each (multi-agent routing).

What running it takes

OpenClaw needs Node 24.16+ or 26.1+ on macOS, Linux or Windows; the installer script provisions Node if it’s missing (install docs). You can also install with npm, the desktop apps, Docker, or a one-click Railway deploy. There are companion apps for iOS and Android.

Where to run it is the real decision:

• Your own laptop is easiest, but the agent stops when the lid closes, and it runs with your user’s access to your files.
• A spare machine or a small server keeps it always on and away from your personal files.
• A cloud deploy is always on without hardware, but your memory and credentials now live on that host.

Expect to use the command line. The maintainers themselves warn that it is risky for people who aren’t comfortable there; a maintainer quoted on Wikipedia put it as “far too dangerous” for anyone who can’t run a command.

What it costs

The software is free. The model is not, unless you run it locally. Each turn sends your message plus memory and tool results to the model, so token use adds up fast with a top-tier model. Notebookcheck reported users and testers spending over $100 in a day.

Ways to keep the bill down: use a cheaper model for routine chat, set a spending limit with your provider, keep scheduled jobs few, or run a local model through Ollama, LM Studio or vLLM (Ollama provider). Local models cost nothing per message but need capable hardware.

The security trade-offs

OpenClaw’s power comes from broad access: your messages, email, files and a shell. That’s also the risk.

• No sandbox by default. Sandboxing is off, so the agent can run commands on the host unless you turn on approvals or the sandbox.
• Prompt injection. Anything the agent reads, such as an email or a web page, can contain instructions aimed at it.
• Third-party skills. Skills are code. In January 2026, Cisco’s AI security team reported a third-party skill that exfiltrated data and performed prompt injection without the user knowing (summarized on Wikipedia).
• Telemetry. A daily update check sends your OpenClaw version, OS, Node.js version and CPU architecture to the project’s server; update.checkOnStart: false turns it off (telemetry docs).

None of this makes OpenClaw unusable, but it does mean the defaults are tuned for capability, not caution. Our guide Is OpenClaw safe? walks through the settings to change.

OpenClaw vs the closest alternatives

The table above compares OpenClaw with three projects people usually consider next:

• Hermes Agent from Nous Research is the nearest rival in size, sends no telemetry by default and can import OpenClaw’s settings, memories and skills. See Hermes Agent vs OpenClaw.
• NanoClaw is a much smaller agent that runs each agent in its own container. See NanoClaw vs OpenClaw.
• NemoClaw isn’t a replacement: it’s NVIDIA’s stack for running OpenClaw itself inside a locked-down sandbox. See what NemoClaw is.

For the full field, see OpenClaw alternatives.