PicoClaw Setup Guide: Go Binary AI Assistant on $10 Hardware

PicoClaw is a Go rewrite of nanobot that compiles down to a single binary, boots in under a second, and uses less than 10MB of RAM. Sipeed, the RISC-V hardware company behind NanoKVM and MaixCAM, built it to run on their $10 LicheeRV-Nano boards. I’ve been running it alongside my nanobot and ZeroClaw setups for the past week, and the resource numbers are real. If you want an AI assistant on hardware that would choke on Python, this is the one to look at.

This guide walks through getting PicoClaw running on a VPS or SBC with MiniMax M2.5 and GLM-5 as your models, Brave Search for web access, and Discord as the chat channel.

If you’re comparing self-hosted bot options, our OpenClaw alternatives roundup includes PicoClaw alongside nanobot, NanoClaw, memU, IronClaw, and ZeroClaw.

What PicoClaw actually is

PicoClaw is an open-source AI assistant from Sipeed. The project started as a nanobot port to Go, and 95% of the migration was reportedly driven by the AI agent itself with human review. The result is a single binary that runs on RISC-V, ARM, and x86 without any runtime dependencies.

The architecture follows the same pattern as nanobot:

You (Discord / Telegram / DingTalk / LINE / QQ / CLI)
    ↓
PicoClaw Gateway (running on your VPS or SBC)
    ↓
LLM Provider (OpenRouter, Zhipu, Anthropic, OpenAI, Gemini, Groq, DeepSeek)
    ↓
Tools (shell, file access, web search, memory, scheduled tasks)

Messages come in from your chat app, PicoClaw routes them to whatever LLM you configured, and the model can use built-in tools to run shell commands, read and write files, and search the web. Everything except the LLM API calls stays on your machine.

How it compares

Why MiniMax M2.5 and GLM-5

Both of these models came out in February 2026 and they work well with a lightweight bot like PicoClaw.

MiniMax M2.5

MiniMax M2.5 is a 230B Mixture-of-Experts model with only 10B active parameters per pass. It runs fast and cheap:

It scores 80.2% on SWE-Bench Verified, matching Claude Opus 4.6 at about 1/20th the cost. The 1M token context window means PicoClaw won’t hit context limits even with long conversations.

GLM-5

GLM-5 from Zhipu AI is a 744B MoE model with 40-44B active parameters:

GLM-5 ranks first among open-source models on BrowseComp (web search agent tasks), so it’s a solid pick for a bot that does a lot of web lookups. Both models are available through their own APIs and through OpenRouter.

Installation

Three ways to get PicoClaw installed. The prebuilt binary is the fastest path.

# Example for Linux amd64 — check releases for latest version
wget https://github.com/sipeed/picoclaw/releases/download/v0.1.1/picoclaw-linux-amd64
chmod +x picoclaw-linux-amd64
sudo mv picoclaw-linux-amd64 /usr/local/bin/picoclaw

wget https://github.com/sipeed/picoclaw/releases/download/v0.1.1/picoclaw-linux-arm64
chmod +x picoclaw-linux-arm64
sudo mv picoclaw-linux-arm64 /usr/local/bin/picoclaw

git clone https://github.com/sipeed/picoclaw.git
cd picoclaw
make deps
make build

make install

make build-all

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

# Set up your config
cp config/config.example.json config/config.json
nano config/config.json

# Start the gateway
docker compose --profile gateway up -d

# Check logs
docker compose logs -f picoclaw-gateway

docker compose run --rm picoclaw-agent -m "What is 2+2?"

After installing, initialize the workspace and config:

picoclaw onboard

This creates the ~/.picoclaw/ directory with a default config.json and a workspace/ folder.

Check that everything’s working:

picoclaw status

Configuring MiniMax M2.5

PicoClaw uses a JSON config file at ~/.picoclaw/config.json. The provider system routes models based on keywords in the model name, similar to how nanobot does it.

Get an API key

1. Go to platform.minimax.io (global) or minimaxi.com (mainland China)
2. Create an account and generate an API key

Add to config

The simplest way is through OpenRouter, which gives you access to MiniMax alongside hundreds of other models:

{
  "agents": {
    "defaults": {
      "model": "minimax/MiniMax-M2.5",
      "max_tokens": 8192,
      "temperature": 0.7
    }
  },
  "providers": {
    "openrouter": {
      "api_key": "sk-or-your-openrouter-key",
      "api_base": "https://openrouter.ai/api/v1"
    }
  }
}

To hit MiniMax’s API directly, configure a dedicated provider:

{
  "agents": {
    "defaults": {
      "model": "MiniMax-M2.5"
    }
  },
  "providers": {
    "minimax": {
      "api_key": "your-minimax-api-key"
    }
  }
}

For the mainland China endpoint, add "api_base": "https://api.minimaxi.com/v1" to the minimax provider.

Test it

picoclaw agent -m "What's 42 * 17?"

If you get a response, MiniMax is working.

Configuring GLM-5 (Zhipu)

PicoClaw has built-in support for Zhipu. The provider routes automatically when it detects glm in the model name.

Get an API key

1. Go to bigmodel.cn
2. Register and create an API key

Add to config

{
  "agents": {
    "defaults": {
      "model": "glm-4.7",
      "max_tokens": 8192,
      "temperature": 0.7,
      "max_tool_iterations": 20
    }
  },
  "providers": {
    "zhipu": {
      "api_key": "your-zhipu-api-key",
      "api_base": "https://open.bigmodel.cn/api/paas/v4"
    }
  }
}

Switching between models

Configure both providers and swap the default model whenever you want. Change "model" to "MiniMax-M2.5" or "glm-4.7" and PicoClaw picks the right provider automatically. No restart needed for CLI use. For the gateway, restart with picoclaw gateway.

Setting up web search

PicoClaw supports both Brave Search and DuckDuckGo. Brave gives better results but needs an API key. DuckDuckGo works out of the box with no key required.

Brave Search

1. Go to brave.com/search/api
2. Sign up for an account
3. The free tier gives you 2,000 searches per month
4. Generate an API key from the dashboard

Add to config

{
  "tools": {
    "web": {
      "brave": {
        "enabled": true,
        "api_key": "your-brave-search-api-key",
        "max_results": 5
      },
      "duckduckgo": {
        "enabled": true,
        "max_results": 5
      }
    }
  }
}

Keep both enabled. If Brave hits a rate limit or fails, PicoClaw falls back to DuckDuckGo automatically. The max_results setting controls how many results get pulled per search. Five is a reasonable default.

Discord setup

Discord is one of the channels PicoClaw supports alongside Telegram, QQ, DingTalk, and LINE.

Create a Discord bot

1. Go to discord.com/developers/applications
2. Click New Application, give it a name
3. Go to Bot in the left sidebar, click Add Bot
4. Copy the bot token

Enable intents

Still in the Bot settings page:

1. Scroll down to Privileged Gateway Intents
2. Enable MESSAGE CONTENT INTENT (required, or the bot can’t read messages)
3. Optionally enable SERVER MEMBERS INTENT if you plan to use allow lists

Get your user ID

1. Open Discord Settings → Advanced → enable Developer Mode
2. Right-click your avatar anywhere in Discord
3. Click Copy User ID

Configure PicoClaw

Add the Discord channel to your ~/.picoclaw/config.json:

{
  "channels": {
    "discord": {
      "enabled": true,
      "token": "YOUR_DISCORD_BOT_TOKEN",
      "allow_from": ["YOUR_USER_ID"]
    }
  }
}

The allow_from array restricts who can talk to the bot. Leave it empty to let anyone in your server use it. For a personal bot, always lock this down to your user ID.

Invite the bot to your server

1. In the Discord developer portal, go to OAuth2 → URL Generator
2. Under Scopes, check bot
3. Under Bot Permissions, check Send Messages and Read Message History
4. Copy the generated URL and open it in your browser
5. Select the server to add the bot to

Start the gateway

picoclaw gateway

Send a message in Discord. The bot should respond. If nothing happens, check picoclaw status and look at the gateway logs.

Full config example

Here’s what a complete ~/.picoclaw/config.json looks like with OpenRouter, Zhipu, Brave Search, DuckDuckGo, and Discord:

{
  "agents": {
    "defaults": {
      "workspace": "~/.picoclaw/workspace",
      "model": "minimax/MiniMax-M2.5",
      "max_tokens": 8192,
      "temperature": 0.7,
      "max_tool_iterations": 20
    }
  },
  "providers": {
    "openrouter": {
      "api_key": "sk-or-your-openrouter-key",
      "api_base": "https://openrouter.ai/api/v1"
    },
    "zhipu": {
      "api_key": "your-zhipu-api-key",
      "api_base": "https://open.bigmodel.cn/api/paas/v4"
    },
    "groq": {
      "api_key": "gsk_your-groq-key"
    }
  },
  "channels": {
    "discord": {
      "enabled": true,
      "token": "YOUR_DISCORD_BOT_TOKEN",
      "allow_from": ["YOUR_USER_ID"]
    },
    "telegram": {
      "enabled": false,
      "token": "",
      "allow_from": []
    }
  },
  "tools": {
    "web": {
      "brave": {
        "enabled": true,
        "api_key": "your-brave-search-api-key",
        "max_results": 5
      },
      "duckduckgo": {
        "enabled": true,
        "max_results": 5
      }
    },
    "cron": {
      "exec_timeout_minutes": 5
    }
  },
  "heartbeat": {
    "enabled": true,
    "interval": 30
  }
}

Config settings explained

Provider system

PicoClaw routes providers by protocol family. OpenAI-compatible endpoints (OpenRouter, Groq, Zhipu) share one code path. Anthropic has its own. Adding a new OpenAI-compatible provider is mostly a config operation with api_base and api_key.

Memory and workspace

PicoClaw stores memory and configuration in the workspace directory:

~/.picoclaw/workspace/
├── sessions/          # Conversation sessions and history
├── memory/            # Long-term memory (MEMORY.md)
├── state/             # Persistent state (last channel, etc.)
├── cron/              # Scheduled jobs database
├── skills/            # Custom skills
├── AGENTS.md          # Agent behavior guide
├── HEARTBEAT.md       # Periodic task prompts
├── IDENTITY.md        # Agent identity
├── SOUL.md            # Agent soul / personality
├── TOOLS.md           # Tool descriptions
└── USER.md            # Your personal info and preferences

Tell the bot to remember something and it writes to memory/MEMORY.md. Sessions are stored per conversation. The bootstrap files (SOUL.md, USER.md, etc.) get loaded into the system prompt every time the bot processes a message.

nano ~/.picoclaw/workspace/USER.md

Add whatever context you want the bot to always have — project details, communication preferences, technical background. This works the same way as nanobot’s workspace files.

Heartbeat and scheduled tasks

PicoClaw can run periodic tasks automatically. Create a HEARTBEAT.md file in your workspace:

# Periodic Tasks

- Check the weather forecast
- Search the web for AI news and summarize

The agent reads this file every 30 minutes (configurable) and runs each task using available tools. For long-running tasks, PicoClaw spawns a subagent that works independently without blocking the main heartbeat loop.

You can also manage one-off and recurring jobs from the CLI:

# Add a reminder
picoclaw cron add --name "morning" --message "Good morning! What's in the news?" --cron "0 9 * * *"

# List all jobs
picoclaw cron list

Jobs are stored in ~/.picoclaw/workspace/cron/ and processed automatically.

Security sandbox

PicoClaw runs in a sandboxed environment by default. With restrict_to_workspace set to true, the agent can only access files and execute commands within the workspace directory.

Protected tools

Dangerous command blocking

Even with restrict_to_workspace set to false, PicoClaw blocks destructive commands:

• rm -rf, del /f, rmdir /s — bulk deletion
• format, mkfs, diskpart — disk formatting
• dd if= — disk imaging
• Writing to /dev/sd[a-z] — direct disk writes
• shutdown, reboot, poweroff — system shutdown
• Fork bombs

The sandbox boundary applies consistently across the main agent, subagents, and heartbeat tasks. There’s no way to bypass it through subagents or scheduled jobs.

Disabling restrictions

If you need the agent to access paths outside the workspace:

{
  "agents": {
    "defaults": {
      "restrict_to_workspace": false
    }
  }
}

Docker deployment

PicoClaw has a Docker Compose setup with separate profiles for agent mode and gateway mode.

Using Docker Compose

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

# Set up config
cp config/config.example.json config/config.json
nano config/config.json

# Start gateway (long-running)
docker compose --profile gateway up -d

# Check logs
docker compose logs -f picoclaw-gateway

# Stop
docker compose --profile gateway down

Agent mode (one-shot queries)

# Ask a question
docker compose run --rm picoclaw-agent -m "What is 2+2?"

# Interactive mode
docker compose run --rm picoclaw-agent

Rebuild after updates

docker compose --profile gateway build --no-cache
docker compose --profile gateway up -d

The compose file mounts config.json as read-only and uses a named volume for the workspace, so your data survives container restarts.

CLI reference

In interactive mode, type exit, quit, or press Ctrl+D to leave.

VPS hosting

PicoClaw runs on just about anything. The single binary with under 10MB RAM means you can deploy it on boards as cheap as $10. For a VPS, a Hetzner CX22 (2 vCPU, 4GB RAM) at €4.35/month is overkill, but it leaves room for other services.

Quick setup on a fresh Ubuntu 24.04 VPS:

ssh root@YOUR_SERVER_IP

# Update system
apt update && apt upgrade -y

# Download binary (check releases for latest version)
wget https://github.com/sipeed/picoclaw/releases/download/v0.1.1/picoclaw-linux-amd64
chmod +x picoclaw-linux-amd64
mv picoclaw-linux-amd64 /usr/local/bin/picoclaw

# Initialize
picoclaw onboard

# Edit config
nano ~/.picoclaw/config.json

# Start gateway in background
nohup picoclaw gateway > /var/log/picoclaw.log 2>&1 &

For a proper daemon setup, create a systemd service:

[Unit]
Description=PicoClaw gateway
After=network.target

[Service]
Type=simple
ExecStart=/usr/local/bin/picoclaw gateway
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target

Save to /etc/systemd/system/picoclaw.service, then:

systemctl daemon-reload
systemctl enable picoclaw
systemctl start picoclaw

Running on cheap hardware

PicoClaw was designed for this. A $10 LicheeRV-Nano with Ethernet or WiFi6 can serve as a minimal home assistant. A $30-50 NanoKVM works for automated server maintenance. A MaixCAM handles smart monitoring use cases.

For old Android phones, install Termux and run the ARM64 binary:

pkg install proot wget
wget https://github.com/sipeed/picoclaw/releases/download/v0.1.1/picoclaw-linux-arm64
chmod +x picoclaw-linux-arm64
termux-chroot ./picoclaw-linux-arm64 onboard

If you want to run local models alongside PicoClaw, check our guide on installing Ollama with Docker. PicoClaw works with any OpenAI-compatible endpoint, so pointing it at a local Ollama server is a one-line config change.

PicoClaw vs nanobot vs ZeroClaw

I run all three at this point, so here’s a direct comparison:

PicoClaw is the easiest to deploy: download a binary, run it, done. nanobot has better channel coverage and a bigger community. ZeroClaw has the most capable memory system and the widest provider support. For the nanobot setup guide, see our full walkthrough. For ZeroClaw, see the setup guide.

If you want to explore other AI coding tools, our AI coding tools comparison covers the current landscape. For MCP basics that work across these assistants, check the MCP introduction for beginners.

This article is also available in Spanish: Guía de Configuración de PicoClaw.