Post History

Current VersionApr 27, 2026 at 03:21

Yep, it's like a teeny tiny OpenClaw, which can run on basically anything with a CPU, and requires basically no prerequisite software infrastructure installed on a machine (no Node, Python, etc.). It's just a tiny binary file that runs on most operating systems including Android, Raspberry Pi, etc., or you can run it on a VPS, your home computer, etc.

All the claw apps are local agents that give 'hands' to an LLM, so that they can work with files and the command line on your hardware, to complete work autonomously. The claw ('agent') applications also provide some sort of interface to interact with the LLM, as in a chat interface - those chat interactions can be via the local command console, an SSH connection to a server (what I typically use), or they enable interactions via platforms like Telegram, Discord, Signal, etc., so that you can text with the agent like you're texting with a human. Need to have the agent clean up some files on your server, or respond to all emails about a given topic? - text it a message to do your bidding.

Most of the claw applications also provide access to skills created by a community, which are basically collections of prompt recipes to use tools to accomplish all sorts of common tasks, to connect with well known 3rd party systems, etc., so that you don't have to go through the time and trouble of testing how to best accomplish those tasks and use those tools (particularly tools enabled by MCP servers).

The claw apps also typically enable some sort of memory/learning capability, so that when you complete tasks, it remembers which approaches were most useful - or even if you just want your agent to remember to call you Bob. Typically, memories are stored in the form of .md (Mardown) files and/or database entries.

Since the claw apps are all just harnesses for LLMs to accomplish tasks autonomously on your own hardware, the choice of LLM is really important. My favorite LLM choice for price/performance is Gemini 3.1 Flash Lite (which I access through an account with Openrouter, but you can choose more expensive LLMs if you have a particularly hard task to complete, or you can use locally hosted LLMs to provide the inference brain for your interactions. Most people tend to wire their claw agents to a remotely hosted commercial LLM API (OpenAI, Anthropic, Google, etc.). Openrouter is generally supported by every agent system, to give you access to virtually every hosted API, for both the commercial frontier models, and the open source models. Definitely get an Openrouter account.

If you're just getting used to working with claw-like agents, I'd suggest Hermes and Pi. Hermes is extremely capable, and Pi is very lightweigh. I think Pi is the best for working with locally hosted LLMs, because it's very capable, and it sends a smaller number of tokens on every interaction with the LLM, which really improves performance.

I use Hermes as the big work horse on my VPSs, to get long challenging development tasks completed. I use Pi for local inference (it's been a god send for working with local LLMs). I think of Nullclaw as a quick little tool to complete smaller tasks on littler machines, such as on my phone, or on some travel netbook, or especially on computers which I might not use again, for example, machines that my parents, friends or clients own, which I don't want to install a lot of software on, just to accomplish a one-off task. It's useful especially when I want a fast, clean environment to install for a special task, for example troubleshooting or configuring other software and/or OS config on a system that I don't work on regularly.

I've also had success getting Nullclaw to perform very long running, deeper, challenging tasks, but at this point I'd certainly choose Hermes or Pi for bigger development projects on my long standing VPS accounts and home servers.

I started a thread about my favorite agent/LLM combinations:

https://aibynick.com/thread/27

I'd focus more on the bigger players like Hermes and PI in the beginning. Openclaw has just seemed clunky, bloated, and inelegant so far to me, but it does have huge community support for existing skills and tools.

Version 2Apr 27, 2026 at 03:21

Yep, it's like a teeny tiny OpenClaw, which can run on basically anything with a CPU, and requires basically no prerequisite software infrastructure installed on a machine (no Node, Python, etc.). It's just a tiny binary file that runs on most operating systems including Android, Raspberry Pi, etc., or you can run it on a VPS, your home computer, etc.

All the claw apps are local agents that give 'hands' to an LLM, so that they can work with files and the command line on your hardware, to complete work autonomously. The claw ('agent') applications also provide some sort of interface to interact with the LLM, as in a chat interface - those chat interactions can be via the local command console, an SSH connection to a server (what I typically use), or they enable interactions via platforms like Telegram, Discord, Signal, etc., so that you can text with the agent like you're texting with a human. Need to have the agent clean up some files on your server, or respond to all emails about a given topic? - text it a message to do your bidding.

Most of the *claw applications also provide access to skills created by a community, which are basically collections of prompt recipes to use tools to accomplish all sorts of common tasks, to connect with well known 3rd party systems, etc., so that you don't have to go through the time and trouble of testing how to best accomplish those tasks and use those tools (particularly tools enabled by MCP servers).

The *claw apps also typically enable some sort of memory/learning capability, so that when you complete tasks, it remembers which approaches were most useful - or even if you just want your agent to remember to call you Bob. Typically, memories are stored in the form of .md (Mardown) files and/or database entries.

Since the claw apps are all just harnesses for LLMs to accomplish tasks autonomously on your own hardware, the choice of LLM is really important. My favorite LLM choice for price/performance is Gemini 3.1 Flash Lite (which I access through an account with Openrouter, but you can choose more expensive LLMs if you have a particularly hard task to complete, or you can use locally hosted LLMs to provide the inference brain for your interactions. Most people tend to wire their claw agents to a remotely hosted commercial LLM API (OpenAI, Anthropic, Google, etc.). Openrouter is generally supported by every agent system, to give you access to virtually every hosted API, for both the commercial frontier models, and the open source models. Definitely get an Openrouter account.

If you're just getting used to working with *claw-like agents, I'd suggest Hermes and Pi. Hermes is extremely capable, and Pi is very lightweigh. I think Pi is the best for working with locally hosted LLMs, because it's very capable, and it sends a smaller number of tokens on every interaction with the LLM, which really improves performance.

I use Hermes as the big work horse on my VPSs, to get long challenging development tasks completed. I use Pi for local inference (it's been a god send for working with local LLMs). I think of Nullclaw as a quick little tool to complete smaller tasks on littler machines, such as on my phone, or on some travel netbook, or especially on computers which I might not use again, for example, machines that my parents, friends or clients own, which I don't want to install a lot of software on, just to accomplish a one-off task. It's useful especially when I want a fast, clean environment to install for a special task, for example troubleshooting or configuring other software and/or OS config on a system that I don't work on regularly.

I've also had success getting Nullclaw to perform very long running, deeper, challenging tasks, but at this point I'd certainly choose Hermes or Pi for bigger development projects on my long standing VPS accounts and home servers.

I started a thread about my favorite agent/LLM combinations:

https://aibynick.com/thread/27

I'd focus more on the bigger players like Hermes and PI in the beginning. Openclaw has just seemed clunky, bloated, and inelegant so far to me, but it does have huge community support for existing skills and tools.

Version 1Apr 27, 2026 at 03:20

Yep, it's like a teeny tiny OpenClaw, which can run on basically anything with a CPU, and requires basically no prerequisite software infrastructure installed on a machine (no Node, Python, etc.). It's just a tiny binary file that runs on most operating systems including Android, Raspberry Pi, etc., or you can run it on a VPS, your home computer, etc.

All the *claw apps are local agents that given 'hands' to an LLM, so that they can work with files and the command line on your hardware. They also provide some sort of interface to interact with the LLM, as in a chat interface - those chat interactions can be via the local command console, an SSH connection to a server (what I typically use), or they enable interactions via platforms like Telegram, Discord, Signal, etc., so that you can text with the agent like you're texting with a human. Need to have the agent clean up some files on your server, or respond to all emails about a given topic? - text it a message to do your bidding.

Most of the *claw applications also provide access to skills created by a community, which are basically collections of prompt recipes to use tools to accomplish all sorts of common tasks, to connect with well known 3rd party systems, etc., so that you don't have to go through the time and trouble of testing how to best accomplish those tasks and use those tools (such as MCP servers).

The *claw apps also enable some sort of memory/learning capability, so that when you complete tasks, it can remember which approaches were most useful - or even if you just want your agent to remember to call you Bob.

Since the *claw apps are all just harnesses for LLMs to accomplish tasks autonomously on your own hardware, the choice of LLM is really important. My favorite for price/performance choice is Gemini 3.1 Flash Lite (which I access through an account with Openrouter (definitely get an Openrouter account)), but you can choose more expensive LLMs if you have a particularly hard task to complete, or you can use locally hosted LLMs to provide the inference brain for your interactions.

If you're just getting used to working with *claw-like agents, I'd suggest Hermes and Pi. Hermes is extremely capable, and Pi is very lightweight - I think Pi is the best for working with locally hosted LLMs, because it's very capable and it sends a smaller number of tokens on every interaction with the LLM, which really improves performance. I use Hermes as the big work horse on my VPSs, to get long challenging development tasks completed. I use Pi for local inference (it's been a god send for working with local LLMs), and I think of Nullclaw as a quick little tool to complete smaller tasks on littler machines, especially when I want a fast, clean environment to install for a special task. I've also had success getting Nullclaw to perform very long running, deeper, challenging tasks, but at this point I'd certainly choose Hermes or Pi for bigger development projects.

I started a thread about my favorite agent/LLM combinations: https://aibynick.com/thread/27

I'd focus more on the bigger players in the beginning.

Previous Versions