In practice it often needs >1GB RAM and a small server or Mac mini, which makes “personal AI agents” surprisingly expensive.

I recently came across PicoClaw, an open-source project by Sipeed that takes a very different approach.

Instead of running large runtimes locally, it acts as a lightweight agent client and delegates reasoning to cloud LLM APIs (GLM/GPT/Claude), while keeping orchestration local.

The interesting part is the footprint:

< 10MB memory usage

< 1s cold start

single self-contained binary

no Node.js or Python

runs on ARM / x86 / RISC-V

So it can run on devices like Raspberry Pi 3B, cheap RISC-V boards (~$10), old Android TV boxes, etc.

Technically it’s rebuilt from scratch in Go, which explains most of the startup and memory improvements. No dependency tree, no runtime environment — just one binary.

Despite the size, it still supports:

shell execution

file operations

web search

speech-to-text

Telegram / Discord / QQ / DingTalk integrations

Quick start is basically:

git clone https://github.com/sipeed/picoclaw.git

cd picoclaw make build ./picoclaw agent

Feels more like a “microkernel” approach to agents compared to heavier stacks.

Interesting direction if you’re experimenting with edge AI or home lab automation.

Repo: https://github.com/sipeed/picoclaw

Site: https://picoclaw.org/