Claw, a single POSIX‑sh script, turns any Linux box into a streaming LLM agent that can run shell commands, manage rolling memory, and switch between OpenAI and Anthropic back‑ends. The installer is a one‑liner: download to /usr/local/bin, make executable, and feed your API key.

Built for lightweight environments, Claw requires only curl + jq and a sh‑compatible shell. It works in Alpine containers, v86 VMs, and low‑cost VPS boxes. When the model outputs <shell>…</shell> blocks, Claw executes them and feeds results back, with an optional confirmation prompt. JSONL logs keep prompts and responses; an LLM compresses overflow into session rules.

A distinctive feature is Mentor mode, where a second assistant reviews and refines the first answer. Named sessions allow isolated histories, and dropping .md files into ~/.config/clawlite/instructions/ injects new context without restarting. Demo clips show Claw listing large log files and auto‑compressing old entries.

With zero dependencies beyond sh, curl, and jq, Claw offers a portable, cost‑effective LLM experience for developers who need agentic shell interaction on modest hardware.