OpenClaw explained: tools, models and the agentic loop

Cedric Clyburn explains OpenClaw through the difference between a chatbot response and software that can act on a request. He traces the agentic loop: assemble conversation history, memory and tool instructions, ask a model what to do, execute a tool call, then return its result to the model until it can answer.

The walkthrough describes OpenClaw as an open source assistant running through a local Node.js service on a laptop, virtual machine or Raspberry Pi. Its gateway handles message routing, sessions and tools, with a UI and CLI for management. Messaging adapters translate requests from services such as Slack and iMessage into a common internal format. The model can run locally or use a hosted API; running the gateway yourself does not mean all processing stays on your machine.

Clyburn also explains browser and terminal tools, Markdown-based skills and scheduled tasks through cron jobs. Rather than loading every skill into the context window, OpenClaw supplies brief metadata so the model can read relevant instructions on demand.

The final section addresses security. Clyburn warns that filesystem access, exposed services and malicious skills can put the host at risk. He describes prompt injection through untrusted emails or webpages and recommends isolated environments and reviewing code and skills before use.