PicoClaw is an open-source, ultra-lightweight personal AI assistant released by Sipeed, a hardware maker based in China. Distributed as a single binary written in Go, its core memory footprint stays under 10MB and it boots in less than a second even on a 0.6GHz single-core CPU. Its defining trait is that it runs on boards costing around $10, such as the Raspberry Pi Zero, or on an old Android phone you no longer use. It works as a self-hosted gateway that lets you call an AI agent from the chat apps you already use — Telegram, Discord, Slack, LINE and more — and since its February 2026 release it has continued to add MCP support and a richer web UI.
Key Features
- Runs in under 10MB of memory: The core memory footprint is below 10MB (recent builds may use roughly 10–20MB). Where comparable self-hosted AI assistants need hundreds of megabytes to a gigabyte, PicoClaw is an order of magnitude smaller. It also starts in under a second, so leaving it running all the time costs almost nothing
- A single binary that runs anywhere: Dependency-free single binaries are provided for RISC-V, ARM, MIPS and x86. The same procedure works on a Raspberry Pi, Android, Docker or an ordinary server
- Connects to 19+ chat apps: Telegram, Discord, Slack, WhatsApp, Matrix, LINE, WeChat, QQ, DingTalk, Feishu/Lark, IRC, MQTT and many other messaging channels are supported. There is no dedicated app to install ─ you talk to the AI from the chat client you already use
- Choose from 30+ LLM providers: Cloud APIs such as OpenAI, Anthropic, Google Gemini, DeepSeek, Qwen, Groq, Mistral, AWS Bedrock and Azure OpenAI, plus local models via Ollama or vLLM. Rule-based model routing can send simple queries to a lighter model
- MCP (Model Context Protocol) support: Connect any MCP server to add tools and data sources. Three transports are supported — stdio, SSE and HTTP — and lazy tool discovery keeps the context window from ballooning as the number of connected servers grows
- Built-in tools and scheduling: Web search (DuckDuckGo is enabled by default and needs no API key), web page fetching, shell command execution, file reading, image loading and text-to-speech ship as standard. Recurring tasks and reminders can be set with cron expressions
- Self-hosted, so your data stays with you: Everything runs on your own device and no cloud service account is required. A configuration-based sandbox and a
.security.ymlfile for isolating and filtering sensitive data are also provided
Pricing
| Plan | Price | Highlights |
|---|---|---|
| Open source | $0 (MIT license) | All features free to use; source code may be modified and redistributed |
PicoClaw itself is entirely free and has no paid tier. The only real costs are usage-based fees for whichever LLM API you connect (OpenAI, Anthropic and so on) and the hardware you run it on. Using a local model through Ollama removes the API cost as well.
Pricing is current as of August 2026. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- Resource use is extremely low, so a spare single-board computer or an old phone can become an always-on AI assistant
- The single-binary distribution makes installation simple ─ no runtime or dependency setup required
- A wide range of supported channels and LLM providers lets you assemble a combination that fits your existing setup
- MIT-licensed open source, so you can read the code to verify its behavior or modify it for your own needs
- Being self-hosted, your data never leaves your machine, which makes privacy easier to reason about
⚠️ Cons
- Setup requires command-line work, editing configuration files and obtaining bot tokens for chat apps, which is a high bar for non-engineers
- The software is free, but practical use still requires either an LLM API subscription or an environment capable of running local models
- The official site lists 512MB of RAM as recommended (64MB minimum); the “under 10MB” figure refers to PicoClaw’s own footprint, and you still need headroom for the OS and networking
- It is still on the 0.x series and under active development. Features arrive quickly, but keeping up with specification changes takes effort
- English-language material is available, but community resources in other languages are still thin
Comparison with Similar Services
| Criteria | PicoClaw | OpenClaw | Nanobot | Ollama |
|---|---|---|---|---|
| Main role | Chat-connected AI agent platform | Full-featured AI assistant platform | Minimal agent implementation | Local LLM runtime |
| Language | Go (single binary) | JavaScript/TypeScript | Python (~4,000 lines) | Go |
| Memory use | Under 10MB (core) | Hundreds of MB to ~1GB | Over 100MB | Model-dependent (GB scale) |
| Chat app integration | 19+ | Broad | Limited | None (API only) |
| License | MIT (free) | Open source | Open source | Open source |
OpenClaw leads on feature coverage, taking an “everything included” approach with browser automation and voice. PicoClaw trims that scope and optimizes instead for lightness and startup speed. Nanobot strips the implementation down even further to favor learning and hacking; PicoClaw was in fact inspired by Nanobot and rebuilt in Go. Ollama runs the LLM itself, so it is not a competitor ─ it pairs well with PicoClaw.
Who Is It For
- People who want to stand up an always-on AI assistant using a Raspberry Pi, a small SBC or an unused Android device
- Those who want to send instructions to an AI from Telegram or Discord without handing their data to a cloud service
- Anyone building small automations for a household or team by combining scheduled runs, reminders and shell commands
- Developers who want to write and connect their own MCP servers to give the agent a bespoke toolset
- Operators who need a 24/7 AI agent with minimal power draw and running cost
Summary
PicoClaw is an open-source project that overturns the assumption that running an AI agent requires a powerful machine. Its sub-10MB footprint and single-binary simplicity turn spare small hardware into a ready-made runtime for an AI assistant. Setup does demand comfort with the command line and configuration files, so it is not for everyone, but for people who are fine with self-hosting and want AI available at all times through a chat app, it is a highly cost-effective option. Start by running it on your PC or in Docker, confirm how it behaves, then move it onto a small board.