Since the advent of OpenClaw, a number of alternatives have sprung up. You can find details on some popular options at ClawCharts. I have been experimenting with some of them, and wanted to capture what I have learned so far.

Why so any options?

Before we get into the details, it’s worth looking into why there are so many options.

To me, there are three main reasons.

1. Regardless of work taking place with the major LLM providers, OpenClaw opened the door to showing how a personal LLM agent could really work for people.
2. ClawdBot itself is just a framework. It is a collection of predefined behaviours that cover how to access different LLM providers, how to communicate with the ClawdBot framework, and predefined LLM skills that give access to local information.
3. With a little technical knowledge (how to open a terminal and run a shell command) it is relatively simple to install. Though this has also resulted in some widely publicised security problems and data leaks when people have opened up their OpenClawd instance to the internet for their own access convenience. It’s worth noting that you have to jump through some fairly technical hoops to do this, as OpenClawd does not allow this in it’s default installation.

So, which Claws have I tried?

Nanobot

Nanobot was one of the earlier projects to be released. The goal of the project is to deliver the same functionality as OpenClaw, with much less code. At time of writing it claims to be doing this using 99% fewer lines of code than OpenClaw. (Around 4,000 SLOC, compared to OpenClaws 430,000.)

I ran Nanobot using Claude as the backend LLM. I wanted to use Matrix as the chat interface, and at the time this was not supported out of the box (it is now), so the first task I gave it was to add support for that communication channel.

Nanobot/Claude created the integration in a few minutes, and less than 10 minutes after the original request I was communicating with the bot via Element (a Matrix client).

Spacebot

Next on my list was Spacebot. At time of testing, the default Quick Start instructions involved pulling the code and building the tool. Using this method I was not able to get the web interface running. Looking around the repo I found another quick start document that started with “Use docker”, and that worked full time.

Like Nanobot, Spacebot did not support Matrix out of the box. I gave it the same task to build support for Matrix, and again I had a comms channel open in Matrix in about 10 minutes.

OpenFang

OpenFang‘s unique selling point is that it is “the agent operating system” pre-configured with a number of skills that are focused on specific tasks, agents (called Hands) that can do some specific tasks, and a number of security systems built in that aim to help OpenFang avoid some of the problems that have plagued OpenClaw. For example, protection is in place to try to prevent prompt injection attacks.

I like OpenFang a lot. But I have made using it difficult for myself by using Z.ai as the backend LLM. Z.ai provides 3x the amount of tokens compared to Claude, at one third of the price. It seems to be a false economy though, as the Z.ai APIs tend to be overloaded more often than Claude, and new requests after hours of not using the service can hit rate limits immediately.

This is a developing story for me at the moment. With OpenFang I have created a virtual company that covers Product Management, Development, Architecture, Compliance, Risk, and Marketing. (Sales and C-Suite is my job 😉 ) The main challenges with getting tasks completed seem to be related to using Z.ai, so I’ll be continuing the experiment with Claude soon.

Others

On my ToDo list I also intend to take a look at:

• Agent Zero
• NanoClaw
• PicoClaw
• IronClaw

What next?

I am very focused on OpenFang at the moment. The way that it handles running teams of agents is the most mature and, notwithstanding Z.ai API stability, has so far delivered some great results and outcomes.

But I’ll also be keeping an eye on what comes out of OpenAI. They recently hired Peter Steinberger, the author of OpenClaw, and it doesn’t take a lot of imagination to understand that they might be the first mainstream frontier model provider to productionise this kind of working out of the box. Anthropic is also working on this, but the ultimate winner will be the expanding choices that are becoming available to use these technologies to produce tangible results.

Whatever happens, it’s an exciting time to be in tech.