Home / Architectures

WTF is Claw?

emergingFirst seen 21 Feb 2026

A personal, always-on agent that runs on your own hardware, takes requests over chat apps, and can schedule its own tasks.

Also called: claws, 🦞

TL;DR

A regular agent waits for you to open a chat window. A claw lives on your machine, is reachable from your usual messaging apps, remembers context, and can do things on a schedule ("every morning, check X and message me"). The name comes from the open-source OpenClaw project and its lobster mascot.

How it works

Characterised by three traits: it runs on personal hardware, it communicates through messaging protocols, and it acts both on direct instructions and on tasks it schedules itself. Andrej Karpathy described claws as a new layer on top of LLM agents that pushes orchestration, scheduling, context, tool calls and persistence further. Because a claw combines private data, inbound messages from outside and the ability to act, it is a textbook case for the lethal trifecta, so permissions and approvals matter.

Origin

Coined / introduced by
Popularised by Andrej Karpathy
When
Feb 2026
Story
Grew out of the OpenClaw project and the many small re-implementations it inspired; Simon Willison wrote it up as a new term of art.

Related jargons

Sources