Lately I’ve been using Pi for all my agentic workflows outside of Claude.
Pi is an open source, third party, command-line AI harness, an alternative to Claude Code and Codex.
Those vendor harnesses are intended to be used with their vendors’ models (Claude and GPT respectively), and they come with many features built-in. Pi is different. You can use it with any model. And it’s minimalist, so it starts with relatively few features. But it’s highly customizable, so you can add what you want.
First impressions: 👍.
It’s great. You should try it
Pi forces you to learn things worth learning
Pi gives me a strong emacs vibe, because it is extremely flexible and customizable. This goes beyond configurable settings. You can easily vibecode extensions which modify its fundamental behavior.
Because of this, I know I can customize Pi to suit me better than any vendor harnesses. Good! However, just like emacs, I know it will cost time and learning to reach that point.
But I suspect this is also good because, right now, Pi sits in a domain of maximally valuable knowledge. Understanding what it does under the hood, and understanding how I might want to produce new workflows, is exactly what is worth learning right now.
For example, with Pi, you need to select the search provider the AI should use. This is a bit annoying, but it is worth understanding how choices about search results affect AI behavior.
More generally, vendor harnesses do not expose the fine details of their automatic prompt engineering — that is, exactly what they put in the context window and send to the model. But Pi gives you visibility and control over exactly that.
Right now, in 2026, this is worth understanding, because we are all still figuring it out and because small improvements in technique can drastically affect how well an AI works.
Maybe in a year, these choices will be understood, optimized, standardized, and blackboxed behind an interface, so you will not need to understand it any more than your CPU’s register count. But for now, this is worth knowing more translates into doing more.
Configurability allows exploring workflows
Pi makes it to add and create new features, via skills or extension. I’ve made one for accounting context window usage and one for replicating Codex’s remote compaction logic. I am actually not much interested in how such extensions are implemented. But the genius of Pi is you just ask the AI to build a feature for you, and Pi is so well-documented and extensible, that it pretty much works.
This flexibility is educational in itself. Emacs is the only (how much other software lets you modify it on the fly, in natural langauge??). But it also gives you the opportunity, when using Pi, to think about the design of a usable agentic harness, to explore and learn how you want to use your tools. This too is valuable.
Qualms
My one qualm about Pi is that I suspect it will require time and learning, not just to improve on the vendor harnesses, but merely to get to parity with them.
Some examples:
- With Pi, you need to configure your own “web_search” skill. What configuration is good enough, or at least as good as what Claude Code provides? Good question.
- Pi compacts the context window into plain text. Codex uses a remote endpoint to compact into encrypted binary blobs. Do they work better? Good question.
- For good agentic task adherence, do I need the “auto-prompting” under the hood, like the task reminder messages which which Claude uses? If so, what’s a reasonable default? Good question.
- With Pi, I have the impression that GPT-5.4’s answers are much longer than they are in Codex. Why? Another good question.
Every one of these good questions is interesting, but also represents time needed just to get caught up to a familiar workflow.
Hackable
However, it is not fair to knock Pi for this.
This situation merely exposes how much the vendor harnesses conceal.
On the other hand, one big benefit of Pi is that it does make it easy to switch across models, which supports refining one’s taste about their strengths.
Like a text editor, an AI harness is so intimately tied to everyday workflow that it should be hackable.
Going to stick with it and see how it goes.