# pi

*Category: AI, terminal | Website: https://buildwithpi.ai/*

a not so shitty coding agent


# pi - the infinitely hackable coding agent

Who would have thought that 2025 would be the year when everyone seems to move to the terminal? At least this is what it feels like if you are in a similar twitter bubble as me. Not that I complain, I only discovered the power of the terminal not too long ago either. The reason for this trend is obviously the avalanche of AI-centric coding tools like Claude Code, which wasn't the first (that title would probably go to [aider](https://aider.chat/) but it was released at a time when LLMs were finally good enough to work on tasks for a decent amount of time with acceptable results. In addition, this was paired with a loss-leader flat-rate subscription in the form of Claude Pro/Max. The plan gave you an allowance of tokens that was enough to work for multiple hours per day, which lead to a hype wave of people using it to build things. Shortly after, OpenAI's answer in the form of codex cli was released, which, in contrast to Claude Code, came with an open source license. Which in my book, is good but not good enough if you are still effectively locked into one provider (It does allow for connecting 3rd party models, but the harness is pretty much tuned to the openai models).

Which is why instead of either one, I actually picked a third party harness in the form of [opencode](https://opencode.ai/) which I have been using ever since and have recommended to many people. I still highly recommend it and I sincerely hope that they will manage to stay independent and will not be acquired (or acquihired) by one of the big AI labs eventually. Because I firmly believe that this kind of tooling needs to be independent and open to work with whichever provider you prefer.

Which brings me to pi, the self-titled 'shitty little coding agent' developed by [Mario Zechner](https://mariozechner.at/) and contributors. What set's pi apart is the extreme minimalism of the base application paired with infinite hackability through extensions. It doesn't support MCP out of the box, because Mario has realized that CLI tools, that can be discovered progressively by the agent > mcp servers that bloat the context window from the get go. Which is what I have been doing myself as well, building [many different cli tools](https://github.com/orgs/byteowlz/repositories) that my agents have access to.

pi is very fast, feels good to use and has an rpc-mode which let's you integrated it into other applications easily. Even though it started as a lean take on ai coding agents, contributors have added many different quality of life features. One very good example is both codex and opencode style message queues: You can either send steering messages which get read by the agent on the very next turn or you can send follow ups, that are only read after the agent has finished implementing the current tasks. It supports most LLM providers already and is also the basis for [Peter Steinberger](https://steipete.me/)'s basement AGI [clawdbot](https://clawd.bot/). And most importantly: pi is unbelievably unpretentious. Mario did't call it shitty little coding agent for nothing. Go give pi a try, you might like it, too!


---

*Tags: coding, TUI, AI*
*Supported OS: linux, macos, windows*