NanoClaw: The Lightweight, Secure OpenClaw Alternative Without the Bloat

BBetter Stack
Computing/SoftwareSmall Business/StartupsInternet Technology

Transcript

00:00:00This is NanoClaw, a very lightweight personal AI assistant
00:00:04coming in at around 500 lines of code
00:00:07and can do things like connect to WhatsApp,
00:00:09read your emails, set recurring tasks and much more
00:00:12using the power of the Clawed Agent SDK
00:00:15and Apple containers if you're on a Mac
00:00:18to keep the whole thing secure.
00:00:19But how does this tiny assistant compare
00:00:21to its bigger, more popular alternative?
00:00:24Hit subscribe and let's get into it.
00:00:25At the time of recording,
00:00:28OpenClaw supports around 15 channels,
00:00:30has more than 45 dependencies and supports many, many skills.
00:00:34NanoClaw, on the other hand, has one channel by default,
00:00:37which is WhatsApp, uses Vercel's agent browser
00:00:40to browse the web and has absolutely no configuration file.
00:00:44Yep, that's right.
00:00:45But its simplicity is what actually makes it so powerful
00:00:49because the code is small and easy enough for agents
00:00:52and even humans to understand.
00:00:53If it doesn't have a feature that you really, really want,
00:00:56you can just ask it to build it
00:00:58and it will do it.
00:00:59I've actually experienced this myself
00:01:00and I'll show you how it works later on in the video.
00:01:03But first, let's go through how to set this up.
00:01:05Actually, let's get the prerequisites out of the way.
00:01:07I highly recommend you get a Clawed subscription
00:01:10if you want to use NanoClaw since it uses the Clawed SDK
00:01:13and that only supports Clawed models by default.
00:01:16This will make more sense in a few seconds.
00:01:18If you're installing this on a Linux machine,
00:01:20make sure you have Docker installed.
00:01:22And if you're using a VPS, do all the standard things
00:01:25like using a non-root user for the installation,
00:01:28disabling public access to SSH
00:01:30and only connecting via tail scale or an equivalent.
00:01:32You'll also need to have Node.js above version 20 installed
00:01:36and have Clawed code
00:01:37because you use a Clawed code slash command
00:01:40to actually set this up.
00:01:41And that's why you need a subscription
00:01:44so you don't waste money burning tokens through API billing.
00:01:47But I'm getting ahead of myself here
00:01:49because you first need to do all these steps
00:01:51before you get to the Clawed code slash command state,
00:01:53which is pretty self-explanatory
00:01:55and connects to WhatsApp using a QR code.
00:01:57From here on, the agent uses Sonnet by default,
00:02:00which you can change of course, but keep that in mind.
00:02:03If you have any issues with the installation,
00:02:05the creator recommends you ask Clawed code directly
00:02:08and it will help you figure out a solution.
00:02:10And also if you want to customise your agent
00:02:12as a slash command to do that,
00:02:14but again, you can also ask it to do so in WhatsApp.
00:02:17You also may need to run the odd NPM command
00:02:20and restart the system a few times.
00:02:22But basically the setup is very simple.
00:02:24And if there's a feature you want
00:02:26that it doesn't say it supports,
00:02:27you can ask it to add that feature.
00:02:30For example, here we can see,
00:02:31I asked the agent if it has the ability to send images
00:02:34and it replies that it doesn't have that ability.
00:02:37So if you scroll down,
00:02:38here I ask it to add the ability to send pictures.
00:02:40It gives me a plan telling me what it's going to change
00:02:43and I give it permission to do so.
00:02:44Now it's added image support by creating a send image tool.
00:02:48So after I restart the service,
00:02:50I ask it to get me a random image from Unsplash.
00:02:53Then after a while, it's able to do that,
00:02:55which we can see over here.
00:02:57And it's able to do that multiple times throughout the day.
00:03:00So I've done it here as well.
00:03:01It's got me an image of an ocean.
00:03:03And I've also done it if we scroll down over here too.
00:03:06Now, if I'm being honest,
00:03:07I wasn't originally sold on NanoClaw.
00:03:10I mean, its setup wasn't as smooth as OpenClaw.
00:03:13It doesn't have a nice dashboard, an identities file,
00:03:16a solve file, and so many things
00:03:18that makes the OpenClaw agent feel so personal.
00:03:22But the bare bones nature of NanoClaw
00:03:25makes it so special
00:03:26because it starts off with a very small feature set
00:03:29and you can ask the agent to add features
00:03:31that are specific to you.
00:03:33Meaning somebody else's NanoClaw
00:03:35will be very different to yours
00:03:36and not just have the features that you want,
00:03:39but the code base will be made to support you
00:03:42as opposed to having something like OpenClaw,
00:03:44which has so many features,
00:03:46but you may only want to have one or two of them.
00:03:48And then you have this code that exists
00:03:49that you'll never use.
00:03:51But its architecture also makes it a bit more secure
00:03:53than OpenClaw.
00:03:54Not just because of its containers
00:03:56that have a five minute timeout,
00:03:57but also because it has a smaller surface area for attacks.
00:04:01A Cisco report revealed that 26% of all agents skills
00:04:05have vulnerabilities.
00:04:06And OpenClaw by default supports a lot of them.
00:04:09Not to mention OpenClaw does not run
00:04:11in a sandboxed environment
00:04:13and has direct access to the system that it's on.
00:04:16Yes, I would have loved Telegram support by default
00:04:19and would love to use models
00:04:21other than the ones from Anthropic.
00:04:23But the creator made this originally for themselves
00:04:26and their own needs, which I completely understand.
00:04:29But the popularity of OpenClaw has introduced a wave
00:04:33of personal AI assistants like Memubot and Nanobot,
00:04:36which is great, but can only lead to one thing.
00:04:40- Skynet has become self-aware.
00:04:41In one hour, it will initiate a massive nuclear attack
00:04:44on its enemy.
00:04:45- What enemy?
00:04:46- Us, humans.

Key Takeaway

NanoClaw provides a secure, minimalist alternative to bloated AI assistants by allowing users to build a bespoke codebase through natural language requests.

Highlights

NanoClaw is an ultra-lightweight AI assistant consisting of only 500 lines of code.

The agent uses the Claude Agent SDK and prioritizes security via sandboxed Apple containers or Docker.

Unlike the feature-heavy OpenClaw, NanoClaw has no configuration files and focuses on a minimal footprint.

Users can expand the assistant's capabilities by simply asking the AI to write and implement new code for itself.

The system requires a Claude subscription and Node.js version 20 or higher for optimal setup.

NanoClaw offers a smaller attack surface compared to alternatives, where 26% of skills often contain vulnerabilities.

Timeline

Introduction to NanoClaw's Philosophy

The speaker introduces NanoClaw as a personal AI assistant that stands out due to its extremely small size of roughly 500 lines of code. It leverages the Claude Agent SDK to perform diverse tasks such as managing WhatsApp messages, reading emails, and setting recurring reminders. Security is a primary focus, utilizing Apple containers on Mac systems to isolate the agent's environment. This section sets the stage by contrasting this 'tiny assistant' with more popular, larger alternatives in the market. It establishes that NanoClaw is designed for users who value speed, simplicity, and safety over pre-packaged complexity.

NanoClaw vs. OpenClaw: Complexity and Customization

The video compares OpenClaw's 15 channels and 45 dependencies against NanoClaw's single default WhatsApp channel and zero configuration files. NanoClaw uses Vercel's agent browser for web navigation, maintaining a code base that is easy for both humans and AI agents to interpret. A unique selling point is the ability to 'ask it to build' any missing features, effectively making the assistant self-evolving. The speaker shares personal experience with this feature, highlighting how the simplicity of the code facilitates rapid modifications. This section emphasizes that while OpenClaw is robust, NanoClaw’s lack of 'bloat' is its true power.

Technical Prerequisites and Setup Process

Setting up NanoClaw requires specific prerequisites including a Claude subscription, Node.js version 20+, and Docker for Linux users. The speaker advises using a non-root user and secure connections like Tailscale for VPS installations to maintain high security standards. The setup process involves a 'claude code' slash command, which helps avoid burning expensive API tokens by utilizing the subscription model. Once active, the agent defaults to the Claude Sonnet model, though this can be customized by the user. Detailed guidance is provided on connecting to WhatsApp via a QR code and utilizing the built-in help commands for troubleshooting.

Live Demonstration: Adding New Features

The speaker demonstrates the self-coding capability by asking the agent to add a feature it initially lacked: sending images. The agent responds by outlining a plan to create a 'send image' tool, which the user then approves through a simple permission prompt. After a quick service restart, the agent successfully fetches and sends random images from Unsplash to the user's WhatsApp. This example illustrates the practical application of NanoClaw's flexible architecture in a real-world scenario. It proves that the assistant is not limited by its initial 500 lines of code, as it can expand its own functionality on demand.

Security Advantages and Final Verdict

The final segment addresses why the speaker eventually chose NanoClaw over OpenClaw despite the latter's superior dashboard and personalization files. A critical security argument is made, citing a Cisco report that 26% of AI agent skills have vulnerabilities, which NanoClaw avoids through a smaller attack surface. Unlike OpenClaw, NanoClaw runs in a sandboxed environment, preventing the AI from having direct, unmonitored access to the host system. While the speaker notes a desire for Telegram support or multi-provider model access, they appreciate the creator's focus on a secure, bespoke experience. The video concludes with a humorous nod to 'Skynet' and the growing wave of personal AI bots like Memubot and Nanobot.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video