Skip to main content
Updated Mar 07, 2026

NanoClaw Hands-On Setup

In Lesson 9, you explored NanoClaw's architecture on paper -- the Body + Brain separation, container isolation, the six-layer reference architecture, and why Layer 3 is the only portable investment. Theory is useful. But you have been using OpenClaw hands-on since Lesson 2, and reading about a different architecture is not the same as touching it. The only way to form a real opinion is to run the same tasks on both platforms and compare what you experience.

That is what this lesson delivers. In the next 40 minutes, you will install NanoClaw, connect WhatsApp, run the same research and file creation tasks from Lesson 3, walk through container isolation in practice, and design the Intelligence Layer for your own profession. You walk away with NanoClaw running on your machine and a written Layer 3 design you can build on for the rest of this book.

macOS and Linux Only

NanoClaw does not support Windows. If you are on Windows, you have two options: use WSL2 (Windows Subsystem for Linux) to run NanoClaw inside a Linux environment, or follow along by reading the outputs while your classmates run the commands. The concepts -- container isolation, platform comparison, Layer 3 design -- apply regardless of which OS you use.


Docker Pre-Flight

NanoClaw runs agents inside Docker containers. Before installing NanoClaw, confirm Docker is installed and running.

docker --version

Output:

Docker version 28.1.1, build 4eba377

If you see command not found or a version error, install Docker first:

  • macOS: Download Docker Desktop and run the installer. Open Docker Desktop once to start the daemon.
  • Linux: Follow the official install guide at docs.docker.com/engine/install for your distribution. Then run sudo systemctl start docker.

Verify Docker is running (not just installed):

docker ps

Output:

CONTAINER ID   IMAGE   COMMAND   CREATED   STATUS   PORTS   NAMES

An empty table is correct -- it means Docker is running with no containers yet. If you see Cannot connect to the Docker daemon, open Docker Desktop (macOS) or start the service (sudo systemctl start docker on Linux).

NanoClaw Defaults to Docker

NanoClaw supports Apple Containers on macOS, but Docker is the default and recommended path. Stick with Docker for this lesson -- it works the same on macOS and Linux, and the commands match what you will see in documentation and community discussions.


Install NanoClaw

Three commands. Under five minutes.

git clone https://github.com/qwibitai/nanoclaw.git
cd nanoclaw
claude

That last command opens Claude Code inside the NanoClaw directory. Claude Code reads the entire codebase -- all ~500 lines of it. This is the 35K token advantage from Lesson 9: Claude Code can load the complete NanoClaw project into its context window, understand every file, and configure the system without you editing a single configuration file.

Now run the setup command:

/setup

Claude Code handles everything: installing dependencies, configuring the WhatsApp connection, setting up Docker containers, and starting the service. Watch what happens -- you are seeing the "agents building agents" pattern from Lesson 9 in action. Claude Code is reading the codebase it can fully understand, following the setup skill instructions, and configuring the system.

Output:

When setup completes, you see a QR code for WhatsApp authentication and a confirmation that NanoClaw is running.

If /setup Asks Questions

Answer based on your preferences. For first-time setup, accept the defaults. If it asks about container runtime, choose Docker.


Connect WhatsApp

The WhatsApp connection follows the same pattern as Lesson 2 -- a QR code scan:

  1. Open WhatsApp on your phone
  2. Go to Settings > Linked Devices > Link a Device
  3. Scan the QR code displayed in your terminal

The terminal confirms the connection. You now have a second AI Employee running on a different architecture, reachable from the same phone.

Separate WhatsApp Number

If you connected your primary WhatsApp number to OpenClaw in Lesson 2, you will need a different number for NanoClaw -- WhatsApp allows only one linked device per external service. Use a second phone number, or disconnect OpenClaw first (openclaw channels logout --channel whatsapp) and reconnect to NanoClaw.


Run the Same Tasks

Open WhatsApp and send NanoClaw the same tasks you ran on OpenClaw in Lesson 3. Running identical tasks on a different platform is the fastest way to understand architectural differences.

Task 1: Research

Research the top 3 competitors in [your industry]. Create a comparison
table with pricing, features, and target market for each.

Use the same industry you used in Lesson 3. When the output arrives, compare it with the competitors.md file your OpenClaw employee created six weeks ago. Notice what is similar (the structure, the agent loop phases) and what differs (response time, tool access patterns, output format).

Task 2: File Creation

Create a file called nanoclaw-test.md with 5 professional goals
for this week, formatted as a markdown checklist.

The agent creates the file -- but where? Inside the container. This is the first concrete difference: NanoClaw's files live inside isolated containers, not on your shared filesystem. Ask the agent:

Where did you save that file? Show me the full path.

The path is inside the container's filesystem. In OpenClaw, files lived on your machine's filesystem alongside everything else. In NanoClaw, each task gets its own isolated space.


See Container Isolation in Practice

Open a new terminal (keep the NanoClaw session running) and check what Docker is doing:

docker ps

Output:

CONTAINER ID   IMAGE          COMMAND   CREATED        STATUS        NAMES
a1b2c3d4e5f6 nanoclaw:dev ... 2 minutes ago Up 2 minutes nanoclaw-agent-...

You see NanoClaw's container running. Each agent task executes inside this boundary. What happens in the container stays in the container:

  • Files created inside the container do not appear on your host filesystem unless explicitly mounted
  • If the agent processes sensitive data (financial records, medical documents), that data never touches your host system
  • When the container is destroyed, all temporary data inside it disappears

This is the architectural inversion from Lesson 9 in practice: nothing accessible unless explicitly granted, compared to OpenClaw's everything accessible unless explicitly restricted.

Run a quick demonstration:

# See what is running inside the container
docker ps --format "table {{.Names}}\t{{.Status}}\t{{.Image}}"

Output:

NAMES                    STATUS         IMAGE
nanoclaw-agent-abc123 Up 3 minutes nanoclaw:dev

One container, one agent session, one isolated boundary. When you ran the research task, the web search, the file creation, and the analysis all happened inside this container. Your host system only received the final response on WhatsApp.


The 35K Token Advantage

Remember from Lesson 9: NanoClaw's core is ~500 lines. Claude Code can load the entire codebase into its context window -- roughly 35,000 tokens including documentation and configuration.

Ask your NanoClaw employee:

How many files are in your codebase? Can you list all of them
with a one-line description of what each does?

NanoClaw responds with a complete map of its own source code. Every file, every purpose, fully understood. Now think about OpenClaw's 430,000+ lines across 52+ modules. No single Claude Code session can hold all of that in context. This is not a judgment of quality -- it is an architectural fact with practical consequences:

  • NanoClaw: Claude Code understands the entire system. The /setup command works because Claude Code can read every file and follow every dependency.
  • OpenClaw: Claude Code can work with individual modules, but cannot hold the full picture. Extensions require understanding specific subsystems, not the whole.

Both approaches have tradeoffs. OpenClaw's size brings features, community, and ecosystem. NanoClaw's size brings auditability, comprehensibility, and the agents-building-agents pattern.


Platform Comparison

You have now used both platforms hands-on. Fill in this comparison from your own experience:

DimensionOpenClaw (Your L02-L08 Experience)NanoClaw (Today)
Setup time~30-45 min (L02)~5 min
Setup methodInstall script + wizardgit clone + Claude Code /setup
WhatsApp connectionQR code scanQR code scan
Where files liveYour filesystem (shared)Inside container (isolated)
Security defaultEverything accessibleNothing accessible
Codebase size430,000+ lines~500 lines
Can Claude Code read the whole thing?NoYes
Community and ecosystem209,000+ stars, ClawHubSmall, growing
Your research task quality[Your assessment][Your assessment]
Your file creation experience[Your assessment][Your assessment]

The last two rows are yours to fill in. Neither platform is universally better -- they serve different stages of business maturity:

StageBest FitWhy
Personal productivityOpenClawRich features, huge ecosystem, fastest path to a working AI Employee
SMB / team deploymentOpenClawCommunity skills, Google Workspace integration, compound workflows
Regulated / enterpriseNanoClaw architectureContainer isolation, full auditability, Programmatic Tool Calling keeps sensitive data in-boundary

OpenClaw gets you from zero to business value in an afternoon. NanoClaw's architecture is what you need when that business handles client data you cannot afford to leak. Most of you will start with OpenClaw and grow into NanoClaw's model as the stakes increase -- and because Layer 3 (your skills and MCP servers) is portable, that transition preserves everything you built.


Design Your Intelligence Layer

This is the capstone exercise for the entire chapter. In Lesson 9, you saw the six-layer reference architecture and learned that Layer 3 (Intelligence) is the only fully platform-independent layer. Now design one.

Choose a profession you know well -- your own, or one you are deeply familiar with. Then complete this design:

Layer 3 Design Template

Profession: ___

Agent Skills (Domain Knowledge -- How to Think):

Skill NameWhat Domain Rules It EncodesWhy an AI Employee Needs This
1.
2.
3.

MCP Servers (Domain Tools -- How to Act):

Server NameWhat System It Connects ToWhat Actions It Enables
1.
2.
3.

Container Isolation Requirements:

  • What data must NEVER leave the container boundary? ___
  • Why? (Regulation, privacy, competitive sensitivity) ___

This exercise is not optional. It is where everything in Chapter 7 comes together: the universal patterns from Lesson 4, the skills from Lesson 5, the security realities from Lesson 5, the Body + Brain architecture from Lesson 9, and the portability standards you just experienced. Your Layer 3 design is the first step toward building an AI Employee for your profession -- and because Skills and MCP are open standards, this design works on any platform.


Try With AI

Prompt 1: Compare Your Experiences

Setup: Use Claude Code or any AI assistant.

I just set up both OpenClaw and NanoClaw and ran the same research
task on each. Here are my observations: [paste your comparison table
from above]. Based on my experience, which platform fits better for
someone in [YOUR PROFESSION]? Consider data sensitivity, setup
complexity, feature needs, and community support.

What you're learning: Platform evaluation as a professional skill. Instead of reading comparison articles, you are generating your own evaluation from hands-on data. The AI helps you weigh tradeoffs you may not have considered, but the experience data is yours.

Prompt 2: Strengthen Your Layer 3 Design

Setup: Use Claude Code or any AI assistant.

Here is my Layer 3 (Intelligence) design for a [YOUR PROFESSION]
AI Employee: [paste your design table]. Review it and suggest:
(1) any critical Agent Skills I missed, (2) any MCP servers that
would be high-value additions, and (3) which skill I should build
first and why.

What you're learning: Iterative design with AI feedback. Your first draft captures your domain expertise. The AI suggests patterns from other domains you may not have considered. Together you converge on a stronger design -- the same bidirectional learning pattern you have been practicing since Lesson 3.

Prompt 3: Container Boundary Analysis

Setup: Use Claude Code or any AI assistant.

For a [YOUR PROFESSION] AI Employee, analyze the data flows in
my Layer 3 design: [paste your MCP servers]. For each MCP server,
classify the data it handles as (a) safe to leave the container,
(b) must stay inside the container, or (c) requires encryption
before leaving. Explain each classification.

What you're learning: Security architecture thinking applied to your own domain. The container boundary is not academic -- it determines whether your AI Employee can legally operate in regulated environments. By classifying your own data flows, you build the instinct to think about data sensitivity before you build, not after a breach.

Flashcards Study Aid