Skip to content

NanoClaw

What it is

NanoClaw is a lightweight, AI-native personal assistant framework designed as a secure, containerized alternative to OpenClaw. It runs on the Claude Agent SDK and prioritizes codebase simplicity and OS-level isolation, fully supporting the MCP 3.0 Task Protocol for reliable tool execution in July 2026.

What problem it solves

It addresses the security risks and code complexity of heavy agent frameworks by providing a minimalist, container-first assistant that evolves through self-modification and composable skills. It ensures that agentic workflows remain secure and private by utilizing FastMCP 3.0 for rapid, type-safe tool discovery and execution.

Where it fits in the stack

Category: Development & Ops / Personal Assistant. It is a lightweight agent runtime for individuals and developers looking for a secure local-first execution environment that integrates seamlessly with Gemma 3 and Claude 4.8.

Typical use cases

  • Secure AI Assistance: Sandboxed task execution for personal local automation.
  • Custom Agent Swarms: Building multi-channel agents (WhatsApp, Telegram, etc.) with strict data isolation.
  • Self-Evolving Skills: Developing agents that modify their own logic through the Anthropic Agent Skills protocol.
  • Local Tool-Calling: Using Gemma 3 with local system tools via the MCP 3.0 Bridge.

Strengths

  • Security-First: Native container isolation; agents run in ephemeral Linux containers by default.
  • Minimalist: Small codebase (under 5k LOC), easy to understand and fork for specific needs.
  • FastMCP 3.0 Integration: Lowest latency for tool registration and execution in the personal assistant category.
  • High Efficiency: Optimized layer templates reduce token costs by up to 40% compared to unoptimized patterns.

Limitations

  • Claude-Centric: Primary optimization is for Claude models, though Gemma 3 support is stable via MCP.
  • Self-Modification Risk: Requires comfort with an assistant that writes its own logic (can be disabled via NC_READONLY_MODE=true).
  • Resource Minimums: Requires at least 4GB RAM and Docker 24+ for the isolation layer to function correctly.

When to use it

  • When you want a personal AI assistant that can be fully understood and customized (low code complexity).
  • When you require strong security via Linux container isolation (Apple Container or Docker).
  • If you prefer a "skills over features" model where the assistant evolves through code transformations.

When not to use it

  • If you require a managed service or a complex, multi-user enterprise framework.
  • If you are not comfortable with an assistant that modifies its own source code to add features.
  • For high-concurrency production workloads that require complex orchestration beyond a single node.

Getting started

Installation

NanoClaw requires Node.js 22+ and Docker.

# Clone the repository
git clone https://github.com/qwibitai/nanoclaw.git
cd nanoclaw

# Install dependencies
npm install

# Initialize with Claude Code
claude /setup

Verifying Isolation

To ensure your agent is correctly sandboxed, run: 

nanoclaw exec "echo 'hello' > /root/secret.txt && ls /root"
Then verify that a subsequent command or a different container instance cannot see that file.

CLI examples

Execution

# Run a one-off task in the sandbox
nanoclaw task "Organize my downloads folder into categories"

# List active agent containers
nanoclaw ps

Skill Management

# Add a new skill via natural language
nanoclaw add-skill "Integrate with my local Obsidian vault"

# View installed skills
nanoclaw list-skills

API examples

Node.js (Agent SDK)

import { NanoClaw } from 'nanoclaw';

const agent = new NanoClaw({
  model: 'claude-4-8-sonnet',
  sandbox: true
});

const response = await agent.run("Summarize README.md and suggest 3 improvements.");
console.log(response);

FastMCP 3.0 Bridge

NanoClaw can bridge local tools to remote agents via FastMCP:

{
  "mcpBridge": {
    "enabled": true,
    "port": 3000,
    "protocol": "fastmcp-3.0",
    "allowedTools": ["filesystem", "bash"]
  }
}

Sources / references

Contribution Metadata

  • Last reviewed: 2026-07-21
  • Confidence: high