PicoClaw Complete Guide — Download, Install, Configure & Build AI Agents on Any Device

Our first PicoClaw article broke down the project from a high-level perspective — what it is, why it matters, and how it fits into the broader AI agent ecosystem. That article became one of our most-read pieces ever.

But the feedback was clear: "How do I actually install this thing? How do I configure it? How do I run it on my Raspberry Pi? Can I use it with llama.cpp? What about Docker?"

This article answers every single one of those questions. This is the definitive PicoClaw installation, configuration, and usage guide — from downloading your first binary to running multi-agent setups with Docker Compose. We're covering every platform, every installation method, every configuration option, and every real-world use case developers are actually building with PicoClaw right now.

If you've ever searched for "picoclaw github", "picoclaw install", "picoclaw termux android", or "picoclaw raspberry pi setup" — this is the only guide you'll need.

What Is PicoClaw — The Complete Picture

PicoClaw is an ultra-lightweight personal AI assistant framework built entirely in Go. Created by Sipeed — a Chinese hardware company known for affordable RISC-V and AI development boards — PicoClaw was designed from the ground up to run AI agent workloads on hardware that costs as little as $10.

The name itself tells you what it's about. "Pico" means tiny. "Claw" references the mantis shrimp (皮皮虾) — a creature that punches far above its weight. That's exactly what this framework does: delivers full AI agent capabilities — task planning, tool execution, web search, code generation, file management — while consuming less than 10MB of RAM.

PicoClaw isn't a chatbot wrapper. It's a genuine AI agent framework — an autonomous system that can decompose complex tasks into steps, select appropriate tools, execute them, store results in persistent memory, and iterate until the job is done. The official tagline says it best: "Tiny, Fast, and Deployable anywhere — automate the mundane, unleash your creativity."

The Origin Story — From OpenClaw to NanoBot to PicoClaw

PicoClaw is the third generation of a remarkable evolution in AI agent architecture:

• OpenClaw (ClawdBot): The original heavyweight. TypeScript-based. 430K+ lines of code. Requires 1GB+ RAM and a Mac Mini ($599+). Think enterprise SUV.
• NanoBot: The Python-based intermediate. Reduced to ~4,000 lines and ~100MB RAM. Better, but still too heavy for edge hardware.
• PicoClaw: The final form. Rewritten from scratch in Go through a self-bootstrapping process where the AI agent itself drove the entire architectural migration. 95% of the core code was agent-generated with human-in-the-loop refinement.

That last detail is extraordinary. The AI agent literally rewrote itself into a more efficient version. This is recursive AI engineering in its purest form.

PicoClaw GitHub — Repository, Stars & Open Source Community

PicoClaw lives at github.com/sipeed/picoclaw . Here's the current state of the repository as of March 2026:

PicoClaw GitHub Releases — Prebuilt Binaries

The releases page provides prebuilt binaries for every major architecture:

• picoclaw-linux-amd64 — For standard PCs, servers, cloud VMs (x86_64)
• picoclaw-linux-arm64 — For Raspberry Pi, Orange Pi, Rock Pi, and Android via Termux
• picoclaw-linux-riscv64 — For RISC-V boards like LicheeRV Nano
• picoclaw-linux-mips — For MIPS-based embedded devices

No Go installation needed. Download, make executable, run. That's the beauty of a single static binary.

Community & Communication Channels

PicoClaw has built a thriving community in just one month:

• Discord: Active community server with dedicated channels for troubleshooting, feature requests, and skill development
• WeChat: Chinese-language community group for Asian developers
• GitHub Discussions: Feature requests, RFC proposals, and weekly community meetings
• X/Twitter: @SipeedIO for announcements

PicoClaw Architecture — How It Actually Works Under the Hood

Understanding PicoClaw's architecture is key to using it effectively. The design follows a local orchestration + cloud reasoning pattern that makes sub-10MB RAM possible.

The Agent Loop

Every PicoClaw interaction follows this cycle:

1. Input: User sends a natural language command (CLI, Discord, Telegram)
2. Planning: The command is sent to a cloud LLM (via OpenRouter, Groq, Ollama, or direct APIs) which decomposes it into actionable steps
3. Tool Selection: PicoClaw matches each step to available tools — file operations, shell commands, web search, code execution
4. Execution: Tools execute locally on the device. File reads, writes, shell commands — all happen on-device
5. Memory: Results are stored in PicoClaw's lightweight persistent memory system under ~/.picoclaw/workspace/
6. Iteration: The agent loops until the task is complete, feeding results back to the LLM for next-step planning

Why This Architecture Works on $10 Hardware

The genius is in what runs where. The heavy computation — language model inference, reasoning, planning — happens in the cloud (or on a separate Ollama server). The lightweight orchestration — the agent loop, tool dispatch, memory management, I/O handling — runs locally. This is why PicoClaw only needs 10MB of RAM. It's not running a language model. It's running the infrastructure that coordinates one.

The Go Advantage

Go was the perfect choice for PicoClaw. Here's why:

• Static compilation: Single binary, no runtime dependencies. No Python virtualenvs, no Node.js, no package managers.
• Cross-compilation: Build for RISC-V, ARM64, MIPS, and x86_64 from any development machine with one command
• Goroutines: Lightweight concurrency for handling multiple tool executions and messaging platform connections simultaneously
• Low memory footprint: Go's runtime is measured in single-digit megabytes, unlike Python (~30MB) or Node.js (~50MB+)

Install PicoClaw — Prebuilt Binaries (ARM64, x86_64, RISC-V)

The fastest way to get PicoClaw running. No Go installation needed. No compilation. Just download and run.

Step 1: Download the Binary for Your Platform

Step 2: Make It Executable

Step 3: Run the Onboarding

The onboard command creates your ~/.picoclaw/ directory with default configuration files. It will prompt you for your LLM API key and preferred provider.

Step 4: Start Using PicoClaw

That's it. Four commands. Under 60 seconds. You're now running a full AI agent on your machine.

Build PicoClaw from Source — Step-by-Step Guide

Building from source gives you the latest features and lets you contribute to the project. Here's the complete process.

Prerequisites

• Go 1.21+ — Download Go
• Git — For cloning the repository
• Make — For the Makefile build system

Install Prerequisites (Ubuntu/Debian)

Clone & Build

Alternative: Direct Go Build

If you prefer not to use Make:

Cross-Compilation

Build for any target architecture from your development machine:

PicoClaw Docker — Container Deployment Guide

PicoClaw ships with Docker Compose configurations for containerized deployment. The Dockerfile uses multi-stage builds with golang:1.25-alpine for minimal image size.

Quick Start with Docker Compose

Docker Compose Services

The default docker-compose.yml includes two services:

• picoclaw-agent: For one-shot agent commands and interactive sessions
• picoclaw-gateway: For running PicoClaw as a persistent messaging bot (Discord, Telegram, IRC)

VPS Deployment Example

Deploy PicoClaw on the cheapest ARM server available (€3.79/month on Hetzner):

PicoClaw Configuration — API Keys, LLM Providers & Models

PicoClaw's configuration lives at ~/.picoclaw/config.json. Here's a complete configuration file with all available options:

{
  "agents": {
    "defaults": {
      "workspace": "~/.picoclaw/workspace",
      "model": "anthropic/claude-opus-4-5",
      "max_tokens": 8192,
      "temperature": 0.7,
      "max_rounds": 50
    }
  },
  "providers": {
    "openrouter": {
      "api_key": "sk-or-v1-xxx"
    },
    "groq": {
      "api_key": "gsk_xxx"
    },
    "ollama": {
      "endpoint": "http://localhost:11434"
    }
  },
  "search": {
    "provider": "tavily",
    "api_key": "tvly-xxx"
  }
}

Supported LLM Providers

Search Provider Configuration

PicoClaw supports multiple web search providers for the agent's research capabilities:

• Tavily: AI-optimized search API (recommended)
• Brave Search: Privacy-focused alternative
• DuckDuckGo: Free, no API key needed

PicoClaw Commands — onboard, agent, gateway

PicoClaw has three core commands. Here's what each one does:

PicoClaw on Raspberry Pi — Complete Setup Guide

Raspberry Pi is the most popular platform for running PicoClaw. Whether you have a Pi Zero, Pi 3, Pi 4, or Pi 5 — PicoClaw runs on all of them. Here's the complete setup.

Step 1: Flash Raspberry Pi OS

Use Raspberry Pi Imager to flash Raspberry Pi OS Lite (64-bit) to your SD card. Lite is sufficient — you don't need a desktop environment.

Step 2: SSH Into Your Pi

Step 3: Download & Install PicoClaw

Step 4: Configure Your API Key

Step 5: Run PicoClaw as a Service (Auto-Start on Boot)

Create a systemd service to run PicoClaw's gateway mode automatically:

[Unit]
Description=PicoClaw AI Agent Gateway
After=network.target

[Service]
Type=simple
User=pi
ExecStart=/usr/local/bin/picoclaw gateway
Restart=always

[Install]
WantedBy=multi-user.target

Your Raspberry Pi is now a 24/7 AI assistant that starts automatically on boot, accessible through Discord or Telegram from anywhere in the world.

PicoClaw on Termux Android — Turn Your Old Phone Into an AI Agent

Got an old Android phone sitting in a drawer? PicoClaw can turn it into a fully functional AI assistant. Here's exactly how.

Step 1: Install Termux

Download Termux from F-Droid (recommended) or Google Play Store. F-Droid has the latest version.

Step 2: Install Dependencies

Step 3: Download PicoClaw ARM64 Binary

Step 4: Run with proot (Required for Termux)

The proot and termux-chroot commands are necessary because Termux's filesystem layout differs from standard Linux. The chroot creates a compatible environment.

Step 5: Configure & Chat

Use PicoClaw with llama.cpp & Ollama — Run Local AI Models

One of the most requested features: running PicoClaw with completely local AI models. No cloud API calls. No API keys. Full privacy. Here's how.

Method 1: PicoClaw + Ollama (Recommended)

Ollama provides an OpenAI-compatible API that PicoClaw can connect to directly.

Step 1: Install Ollama

Step 2: Pull a Model with Tool Support

Step 3: Configure PicoClaw for Ollama

{
  "agents": {
    "defaults": {
      "model": "glm4"
    }
  },
  "providers": {
    "ollama": {
      "endpoint": "http://localhost:11434"
    }
  }
}

Method 2: PicoClaw + llama.cpp (Advanced)

For developers who want maximum control, you can run llama.cpp directly and point PicoClaw at its OpenAI-compatible API server.

Build & Run llama.cpp Server

"providers": {
  "ollama": {
    "endpoint": "http://localhost:8080"
  }
}

PicoClaw Skills System — Extending Agent Capabilities

PicoClaw's skill system is how you extend the agent's capabilities beyond its built-in tools. Skills are modular, self-contained packages that add specialized knowledge and capabilities.

What Is a PicoClaw Skill?

Each skill is a directory containing a SKILL.md file that defines:

• Name: Unique identifier for the skill
• Description: What the skill does (used by the agent for skill selection)
• Instructions: Detailed guidance for the agent on how to use the skill
• Scripts: Optional shell scripts or binaries the skill can invoke
• References: Documentation and examples
• Assets: Supporting files the skill needs

Built-in Skills

PicoClaw ships with several built-in skills in the skills/ directory:

• skill-creator: Create or update other skills — the meta-skill for extending PicoClaw itself
• web-search: Internet research and information gathering
• code-execution: Write and execute code in the agent's environment
• file-manager: Create, read, update, and delete files
• shell-executor: Run shell commands on the host device

Creating a Custom Skill

You can create custom skills by adding a new directory under ~/.picoclaw/workspace/skills/. The agent can even create skills for you:

PicoClaw will use the skill-creator meta-skill to generate a complete skill package with SKILL.md, scripts, and configuration.

PicoClaw Dashboard — Monitoring & Management

While PicoClaw is primarily a CLI tool, the community has been building dashboard interfaces for monitoring and management. The official roadmap includes a built-in web dashboard, but several community solutions already exist.

Current Dashboard Options

• Discord-based Dashboard: Use PicoClaw's gateway mode with Discord for a visual chat-based interface with rich embeds showing task status, memory contents, and tool execution logs
• Telegram Bot Interface: Full conversational UI with inline buttons for common commands
• Community Web Dashboards: Several open-source web UIs have been built by community members, available through GitHub searches

Monitoring PicoClaw

You can monitor PicoClaw's resource usage and agent activity through:

PicoClaw Multi-Agent — Collaborative AI Framework

One of PicoClaw's most ambitious roadmap features is the multi-agent collaboration framework — multiple PicoClaw instances working together on shared tasks with shared context.

How Multi-Agent PicoClaw Works

The concept is straightforward but powerful:

• Agent Specialization: Each PicoClaw instance is configured with different skills and roles — one handles web research, another handles code execution, a third manages file operations
• Shared Workspace: All agents share a common workspace directory (via networked filesystem or shared volume)
• Task Decomposition: A coordinator agent breaks complex tasks into subtasks and delegates to specialized agents
• Result Aggregation: Outputs from all agents are collected and synthesized

Docker Compose Multi-Agent Setup

Running multiple PicoClaw agents is natural with Docker Compose:

services:
  researcher:
    image: picoclaw:latest
    environment:
      - PICOCLAW_ROLE=researcher
    volumes:
      - shared-workspace:/workspace
  
  coder:
    image: picoclaw:latest
    environment:
      - PICOCLAW_ROLE=coder
    volumes:
      - shared-workspace:/workspace

  coordinator:
    image: picoclaw:latest
    environment:
      - PICOCLAW_ROLE=coordinator
    volumes:
      - shared-workspace:/workspace

volumes:
  shared-workspace:

Imagine a fleet of $10 boards, each running a specialized PicoClaw agent, collaborating on complex automation tasks. A research agent gathers data. A coding agent processes it. A reporting agent generates outputs. All orchestrated automatically. This is the near future of edge AI.

PicoClaw Trading Bot — Automated Trading Use Case

One of the most talked-about community use cases is building a PicoClaw-powered trading bot. Here's the concept and how developers are approaching it.

How a PicoClaw Trading Bot Works

Developers are building trading assistants (not autonomous traders) using PicoClaw's agent capabilities:

• Market Research: PicoClaw's web search skills gather real-time market data, news, and sentiment analysis
• Data Processing: The agent processes and summarizes market information into actionable insights
• Alert Generation: Based on configurable conditions, PicoClaw sends alerts via Discord or Telegram
• Report Generation: Automated daily/weekly market summary reports

The key advantage: PicoClaw runs 24/7 on minimal hardware, continuously monitoring markets and providing analysis at near-zero infrastructure cost.

PicoClaw Use Cases — From IoT to DevOps to Education

1. Personal Home AI Assistant

Deploy PicoClaw on a Raspberry Pi connected to your home network. Ask it to monitor emails, summarize RSS feeds, manage smart home devices, draft responses, and schedule reminders. Always on, always ready, for under $50 in hardware.

2. IoT Edge AI Agent

Embed PicoClaw into IoT deployments — smart agriculture sensors, industrial monitoring, warehouse automation. The agent processes sensor data locally, makes decisions, and only reaches out to cloud APIs when complex reasoning is needed.

3. DevOps & Server Monitoring

PicoClaw on NanoKVM devices monitors servers, responds to alerts, executes maintenance scripts, and troubleshoots issues. For small teams managing multiple servers, this is transformative.

4. Developer Productivity Assistant

Run PicoClaw as a full-stack engineering assistant. It reads your codebase, writes code, manages files, searches documentation, and executes shell commands. Like having a junior developer that works 24/7.

5. Smart Surveillance & Monitoring

On camera-equipped boards like MaixCAM, PicoClaw performs person detection, anomaly monitoring, and intelligent alerting — all on-device with AI-powered analysis.

6. Education & Research

At $10 per deployment, PicoClaw is perfect for teaching AI agent concepts. Universities can give every student their own AI agent to experiment with — financially impossible with traditional setups.

7. Content Automation

Use PicoClaw for automated content generation, social media management, email drafting, and research compilation. The generative AI content tools space is where PicoClaw shines as a lightweight automation engine.

PicoClaw vs OpenClaw — Complete Performance Comparison

This is the comparison everyone searches for. Here's the definitive, data-backed breakdown.

When to Choose PicoClaw

• Edge AI deployment on minimal hardware
• Always-on personal AI assistants at near-zero cost
• IoT and embedded AI applications
• Resource-constrained environments
• Simple, fast deployment without dependency management

When to Choose OpenClaw

• Complex multi-step enterprise automation
• Production-grade deployments requiring stability
• Advanced plugin ecosystem with mature integrations
• Teams with Mac/Linux workstations and adequate resources

Read our complete OpenClaw deep dive for the full picture on OpenClaw's capabilities.

Sipeed PicoClaw Roadmap — What's Coming Next

Sipeed has published a detailed community roadmap outlining PicoClaw's future direction. Here are the key priorities:

🚀 Core Optimization: Extreme Lightweight

The team is committed to maintaining the sub-10MB footprint even as features expand. Recent PRs have temporarily increased usage to 10–20MB, and optimization is the top priority once the feature set stabilizes.

🔒 Security Hardening

PicoClaw explicitly warns against production deployment before v1.0 due to unresolved security considerations. The security roadmap includes sandboxed tool execution, API key encryption, and network isolation.

🤝 Multi-Agent Collaboration

The planned multi-agent framework will allow multiple PicoClaw instances to collaborate on shared tasks — the most requested community feature.

📱 Platform Integrations

Beyond the current Discord support, the roadmap includes Telegram, Slack, WeChat, and IRC integrations. IRC support was just merged in March 2026.

🎨 Community Logo

The team has opened Issue #297 for community logo submissions. They're looking for a cute mantis shrimp design — reflecting PicoClaw's mascot and the Chinese meme "皮皮虾，我们走！" (Mantis shrimp, let's go!).

🧩 Skill Marketplace

A community-driven skill marketplace where developers can share and install custom skills — similar to VS Code extensions or npm packages.

PicoClaw Free Availability — Is It Really Free?

Yes. PicoClaw is completely free and open source under the MIT license. You can download, modify, distribute, and even sell products built with PicoClaw without restrictions.

What's Free

• ✅ PicoClaw software — 100% free, MIT licensed
• ✅ All source code on GitHub
• ✅ Prebuilt binaries for all architectures
• ✅ Docker configurations
• ✅ Community support (Discord, GitHub)
• ✅ Skills system and built-in skills

What May Cost Money

• 💰 LLM API calls: Cloud LLM providers charge per token. However, Groq offers a generous free tier, and Ollama is completely free (local models)
• 💰 Hardware: You need some device to run PicoClaw on (as low as $10)
• 💰 Search API: Some search providers charge per query (DuckDuckGo is free)

Zero-Cost Setup

It's entirely possible to run PicoClaw at zero cost: use an old Android phone (free), Ollama with a local model (free), and DuckDuckGo for search (free). Total cost: $0.

PicoClaw as a Golang AI Agent Framework

PicoClaw is arguably the most significant Golang AI agent framework to emerge in 2026. While Python dominates the AI ecosystem, PicoClaw proves that Go is uniquely suited for agent infrastructure.

Why Go for AI Agents?

• Performance: Go's compiled binaries are orders of magnitude faster than Python scripts for I/O-heavy agent operations
• Concurrency: Goroutines handle multiple tool executions, API calls, and messaging connections with minimal overhead
• Deployment: Single static binary eliminates "works on my machine" problems. No virtualenvs, no package conflicts
• Cross-platform: One codebase compiles to every major architecture including RISC-V
• Memory: Go's garbage collector and runtime are lean enough for embedded systems

For developers building AI agent-powered business assistants, PicoClaw demonstrates that the Go ecosystem is mature enough for production AI agent development.

PicoClaw Board — Hardware That Powers It

Sipeed manufactures the hardware that PicoClaw is designed for. Here are the officially tested boards: