Post History

I gave up on Claude Code a while back because it's intimately tied to the Anthropic LLMs, and those models are way overpriced for what they do.

Openrouter provides some special 'Exacto' versions of many of the best open source models, but they still didn't do a very good job in Claude Code.

I switched one agentic end-user project from Claude Code to Goose, and it instantly did better work than Claude Code did with the open source models (and all in all I was able to get the project more easily completed with Goose).

OpenClaw made an obviously huge splash in the world of AI agents recently, but it's not really so focused on software development (it's intended to function more like a general personal assistant).

I've come to prefer Pi, which is the lightweight core that OpenClaw was built on. It has four simple main tools: Read, Write, Edit, and Bash, which give it a lot of power, while helping it stay minimal. It's optimized enable LLMs to write and edit code, it sends far fewer tokens to the LLM, and it users fewer other system resources.

Pi doesn't have all the connections to messaging apps built in, like Openclaw does, it doesn't have the big skills ecosystem that OpenClaw has, and there are other features such as the permission system which were added to OpenClaw, that are not in Pi, but Pi is lightweight and very capable. It's basically meant to be a command line tool which can do development work and complete complex long tasks, for very little money, especially paired with Gemini 3.1 Flash Lite Preview (that is the best bang for the buck hosted LLM I've found yet).

FUTURE EDIT: use deepseek-pro-4 for most challenging tasks. Use Gemini 3.1 Flash Lite for utility tasks. Gemini is fast and intelligent enough for many types of agentic work, but deepseek is much smarter for development work, and the output is less expensive.

It just takes a few minutes to give it a try, and to really start using it. Install it with:

npm install -g @mariozechner/pi-coding-agent

Set your API in an environment variable:

export OPENROUTER_API_KEY=

Run 'pi'

type /model to set your model.

After that you can just ask it how to use it, and how to do anything.

See more about it at:

https://pi.dev

I gave up on Claude Code a while back because it's intimately tied to the Anthropic LLMs, and those models are way overpriced for what they do.

Openrouter provides some special 'Exacto' versions of many of the best open source models, but they still didn't do a very good job in Claude Code.

OpenClaw made an obviously a huge splash in the world of AI agents recently, but it's not really so focused on software development (it's intended to function more like a general personal assistant).

I've come to prefer Pi, which is the the extremely lightweight core that OpenClaw was built on. It has four simple main tools: Read, Write, Edit, and Bash, which give it a lot of power, while helping it stay minimal. It's optimized enable LLMs to write and edit code, it sends far fewer tokens to the LLM, and it users fewer other system resources.

Pi doesn't have all the connections to messaging apps built in, like Openclaw does, it doesn't have the big skills ecosystem that OpenClaw has, and there are other features such as the permission system which were added to OpenClaw, that are not in Pi, but Pi is lightweight and very capable. It's basically meant to be a command line tool which can do fantastic development work, and complete complex long tasks, for very little money, especially paired with Gemini 3.1 Flash Lite Preview (that is the best bang for the buck hosted LLM I've found yet).

It just takes a few minutes to give it a try, and to really start using it. Install it with:

npm install -g @mariozechner/pi-coding-agent

Set your API in an environment variable:

export OPENROUTER_API_KEY=

Run 'pi'

type /model to set your model.

After that you can just ask it how to use it, and how to do anything.

See more about it at:

https://pi.dev

Previous Versions