You know that drawer of dead electronics everyone has. A Raspberry Pi that never became a weather station, an old Android phone with a cracked corner, a tiny board someone gave you at a conference. Most AI assistants won't touch hardware like that. They want a beefy machine, a gigabyte of memory just to sit idle, and a boot time you can go make coffee during. PicoClaw looks at that drawer and sees a home.
PicoClaw is an open-source personal AI assistant from the hardware company Sipeed, and its whole pitch is smallness. It runs on a board that costs about ten dollars, uses less than 10MB of memory at its core, and starts up in under a second. That's the entire trick, and it's a good one.
Quick facts
License: MIT (free and fully open source). Language: Go. Last active: actively developed, with v0.2.9 shipping in May 2026 and roughly 29,500 GitHub stars as of 4 Jul 2026. Still pre-1.0, so treat it as early software.
What it is
An AI agent is a program that takes your plain-language request, decides which steps to take, and carries them out. It might answer a question, run a command, search the web, or remember something for later. PicoClaw is one of those, except it was rebuilt from scratch in Go, a programming language known for producing small, fast single-file programs. The result is a tool that behaves like the assistants you'd run on a laptop, packed into something that fits on a credit-card-sized computer.
The project also leans on a couple of modern standards so it isn't locked to one brain. It supports the Model Context Protocol (MCP for short, a common way for AI tools to connect to outside services), and it can talk to more than 30 different language-model providers, from OpenAI and Anthropic to a model running locally on your own machine through Ollama. You pick the brain; PicoClaw handles the body.
The real-world problem it solves
Say you want a little always-on helper. Something that watches a folder, pings you on Telegram when a job finishes, or answers questions from a chat app without your main computer being awake. Doing that with a heavyweight assistant means leaving a real machine running, and a Mac mini or a full server pulling power around the clock for one small job feels silly.
PicoClaw fits on a board you can leave plugged into a wall socket and forget about. It sips power, boots instantly if it ever restarts, and doesn't need a screen or a keyboard. The nine-dollar LicheeRV-Nano the team keeps pointing to is smaller than a matchbox.
Where it's useful
- A tinkerer with a LicheeRV-Nano in a drawer turns it into a home helper that answers questions over Telegram and quietly logs the day's reminders, all from a board that cost less than lunch.
- A sysadmin drops PicoClaw onto a NanoKVM device already attached to a server, then talks to it through Slack to check logs or kick off routine maintenance without opening a laptop.
- A hobbyist wiring up a MaixCAM smart camera uses PicoClaw to react when it spots a person at the door, sending a heads-up to a phone instead of writing camera glue code by hand.
- Someone with a decade-old Android phone gathering dust installs the PicoClaw APK and gives the handset a second life as a pocket AI assistant, no rooting or terminal wizardry required.
Why it stands out
Two things make PicoClaw genuinely interesting. First, the numbers are real and stark: under 10MB of memory versus the gigabyte-plus that similar assistants expect, and a sub-second boot on a processor running well below a single gigahertz. When a tool is this light, where you can run it changes entirely.
Second, the way it was built is a story in itself. The team says about 95% of the core code was written by an AI agent and then refined by humans reviewing the output, a "self-bootstrapping" process where the assistant helped port and shrink its own code. Whatever you think of that approach, the finished binary runs across RISC-V, ARM, MIPS, and x86 from a single build, which is not a small feat.
One honest caveat: the project is moving fast, and the maintainers openly note that recent builds sometimes creep up to 10-20MB of RAM as new features land, with cleanup planned later. They also warn it isn't ready for production before v1.0. This is a fun, capable thing to play with, not something to bet a business on yet.
How to get started
The easiest path is the official site, picoclaw.io, which detects your platform and hands you a one-click download. If you'd rather grab a binary or build it yourself, the repo has releases and a Makefile. A minimal command-line setup looks like this:
picoclaw onboard # creates the config and workspace
picoclaw agent -m "What is 2+2?" # one-shot question
picoclaw agent # interactive chat
picoclaw gateway # connect it to a chat app
You add an API key for whichever model provider you want, optionally connect a channel like Telegram or Discord, and start talking. There's also a browser-based launcher for people who'd rather not touch a terminal at all.
Takeaway
PicoClaw is a neat answer to a question most AI tools ignore: what if the assistant had to be tiny. It won't replace a full desktop agent, and being pre-1.0 it has rough edges, but the idea of a capable helper living on a ten-dollar board in the corner of a room is legitimately fun. If you have spare hardware and an afternoon, it's an easy one to try.