Hacker News Reader

Comment by inatreecrown2 | Hacker News Reader

I read most of the post, went to the linked project and still don't have an idea what this is about.

Pi is a coding agent harness, like Claude Code, but significantly better and more elegant. Here is a good post describing it: https://lucumr.pocoo.org/2026/1/31/pi/

I switched a few months ago and have not looked back. Unfortunately Anthropic blocked access from Claude Subscription users today, but that's a different story.

I switched about a month ago, looked back once for about 10 minutes and decided I'm officially done w CC. I didn't realize what a dull knife CC is until I tried a really sharp one, and that's Pi.

Can you elaborate? I want a maximalist setup. I like that CC and Codex are maximalist. If I install Pi, I am going to end up using oh-my-pi and installing a trillion plugins to get a Claude Code-like experience (or better/more feature-heavy). Is there any point in me even trying Pi, or should I just stick with Claude Code?

> I want a maximalist setup.

> Is there any point in me even trying Pi, or should I just stick with Claude Code?

I'd say stick with Claude Code. I'm a minimalist and hate everything about Claude Code!

Sorry I'm late but stick with CC. I introduced a coworker to Pi and spent most of the morning feeling like I should apologize for it not doing this or that out of the box.

Makes sense

I don't know much about the factors that determine why one AI coding harness is better than others. Is it system prompts? Or just personal preference in terms of the UX and there isn't actually a better output between using CC or Pi?

So what makes Pi better than CC? Is it better than OpenCode?

There's a recentish YouTube talk when he introduces the concept and contrasts against those.

My (oversimplified) summary: it's like vim versus an IDE. Good for tinkerers and obsessives who like small, sharp and customisable tools.

> vim versus an IDE is exactly how I describe it to some of my coworkers who are old enough to have used vim.

CC is a steaming pile of vibe-coded bloat. If that rocks your boat, go knock yourself out.

Is it better than OpenCode? It's certainly much smaller and doesn't have a client-server architecture - already that is a big win.

OpenCode is pretty terrible imo. Not very privacy minded.

I've been considering playing around with opencode. Can you expand on what you mean by this?

The harness or the tool is ok but all the defaults as part of the paid pieces of the tool have really bad privacy decisions. So they offer Zen as a pay as you use credit system with access to the models they think work best with the harness. Their own stealth model in it along with a number of the leading new models are always-on sharing data for training purposes. They don’t make this immediately obvious either you have to click through links on their website to see the breakdown.

I am not usually super privacy minded but if you already made it nonobvious this is happening I don’t really trust the underlying tool.

https://opencode.ai/docs/zen/#privacy

Above is the link. The front page says your privacy is important and says they don’t do training on your data except for the following exceptions which links to this page. Then even their own model is training on your data except there is no opt out. So if you pay for zen and you select one of these models in the tool you have no clue it’s auto training on your data.

Gotcha, thanks. That is a bit sus

You mention the client-server architecture of opencode. Is that a local server or is it calling home to opencode servers?

One of the ideas I like about opencode is the ability to prompt and such from a web browser. So I'm curious if that is the client-server architecture you are talking about, or if it's something else.

For reference, I used replit for some vibe-ish coding for a little bit and really liked that I could easily prompt and view output on my phone when hanging out away from my computer. Or while waiting at the airport for example.

(RIP OG replit by the way. They've pretty much completely pivoted from a REPL to AI, which is pretty hilarious to me given the company name xD)

My experience with harnesses is entirely about UX, personally. You could just use an LLM directly and pipe its output directly into your source files, but that would produce terrible results in practice. Harnesses / agents are just better versions of “curl https://llm.com > source.{py/js/cpp/etc}”, imo

Long term I’m bullish on an open source harness “winning” the foot race, in a similar way that Linux “won” over Windows and MacOS (that is, debatably)

Which model are you using it with? I imagine it's bring your own model?

To what extent do you feel the harness contributes relative to the model?

To put another way, how much inferior can the model be with a superior harness to achieve a similar result?

Pi doesn't claim to get better results. It is conceptually simpler, smaller, and more transparent to the end user than most harnesses. It's as much about the things it doesn't do as about what it can do.

Significantly! See this recent post „Compare harnesses not models: Blitzy vs GPT-5.4 on SWE-Bench Pro” https://quesma.com/blog/verifying-blitzy-swe-bench-pro/

I've been using oh-my-pi, for the simple (and possibly naive/stubborn) reason that it doesn't try to get me to install it as global npm dependency in /usr.

I am not a web developer, I don't need npm and I don't want it clobbering my /usr (which is immutable on many modern distro's anyway). Doesn't exactly inspire confidence in the project to me.

oh-my-pi's installer installs a bun bundled binary in my users .local folder. That's much more user-friendly.

pnpm installs to ~/.local as well

They're friendly for the user audience that doesn't care about these things. The location is a minor issue compared to many of the capabilities they come with. For the slightly more tech savvy, they should really be running these harnesses in a contained environment with net cap dropped, for instance.

You can add that via an extension: https://github.com/carderne/pi-sandbox

The price of flexibility is, pi is not opinionated about adding sandboxing out-of-the-box, it gives you options on how you want to do it. You either do it with linux containers, with a dedicated VM, or just bubblewrap. It is nice that it gives you a way to hook into it in a very easy way though.

I use mise to install it in a manageadble way (and basically everything else on my computer).

You can get npm to install into `~/.local`.

Put this into `~/.npmrc`:

```

prefix = ~/.local

```

Bam, `npm install -g` contstrained to your `~/.local`

> significantly better and more elegant.

Can you please give us concrete reasons that make Pi better than Claude Code for you? And what model are you using for it since Claude is not available.

> Can you please give us concrete reasons that make Pi better than Claude Code for you?

I'm not the person you're replying to, but I value good architecture and elegance for aesthetic and maintainability reasons, the MIT license for future-proofness.

Claude Code has been steadily getting worse, it's a steaming pile of vibe-coded bloat and I don't foresee good future for the CC agent. Why would I use it?

There are many people who choose their tools using more criteria than just "how good is the immediate output now".

Now which is the best alternative?

I think it has something to do with a core library that is or was used in one of the other AI-related tools that got some buzz recently.

I'm not sure if this library or if the AI tool itself really matter all that much, but a lot of people seem to think that they do, or will anyways, which is probably why this post has no context. Because the author assumes that both the library and the tool are well known, but I think they're only well known by those in the AI world.

You know how it is. You only know what you know, and it's hard to remember what it was like when you didn't know. When you're around people who talk about the same things all the time, it's easy to just assume that's what's everyone's talking about.

There are so many of these little tools coming and going these days it's hard to keep track. Some of them might end up mattering, but most of them probably won't.

Mario Zechner wrote Pi, an agent framework, and wrote Pi (yes the names are confusing), a coding harness (like Claude Code) on top of it. OpenClaw uses Pi the framework, so now Mario Zechner is joining Armin Ronacher's company.

Taken verbatim "Maybe you've heard about the little app called OpenClaw. OpenClaw is powered by pi. That made me collateral of Peter's success. Especially after Armin thought it's a good idea to tell the whole world about the relationship between OpenClaw and pi on his blog."

Indeed. Also what a very annoying website.

Looks like LLM-generated to me

This could be one of my least favorite website styles ever. Too content dense is overwhelming, but this went WAY too far in the other direction.

If I can't tell what your company is/does without clicking/interacting with the site, what are you even doing?

My other least favorite style is where stuff flies in from the side as you scroll down.

There are some people that are "way too deep" into what they think is the "AI ecosystem". The reality is that no such thing exists and most people opening Claude daily have never heard of their harness, tool, stack or whatever. It has about half the relevancy of "MCP", which is such a similar concept where hanger ons try to catch a steam train.

Wow, I couldn't disagree more!

Claude Code is a steaming pile of vibe-coded bloat.

Pi has been elegantly designed with a minimalist mindset.

> It has about half the relevancy of "MCP", which is such a similar concept where hanger ons try to catch a steam train.

I'm not even sure what you're saying here. Are you saying MCP and Pi are similar concepts? Or that both are hanger-ons? I believe the steam train of Claude Code is heading for a wall and Pi has no intention to catch it.

> Claude Code is a steaming pile of vibe-coded bloat.

Not disagreeing but I'm curious how much it really matters.

I use OpenCode which is similarly a steaming pile of vibe-coded bloat and get pretty good results!

I'm open to switching to pi, which I've read about occasionally on HN, but I'm struggling to see where it might have practical benefits versus OpenCode.

The idea that pi is now under corporate governance with a potentially sustainable business model does intrigue me, but there's risk in adopting "open core" software too due to the potential rug pull and incentives to gate important features behind more restrictive licenses.

I think they are similar to Cursor. They are fads. You need to make the model, not smother it with your ingenuity in wrapping it.

One of the selling points of Pi is that it doesn't "smother the model with ingenuity": it's a minimal harness with a focus on transparency, includes a minimal system prompt, and does not add needless tokens to what you are sending to the model.

What agent do you use then? To me it seems most are terrible, incl Claude Code, OpenCode, and the other usual ones. I never felt an urge to use one. At best I used one very reluctantly as a magical black box.

Pi does seem special, at least to me. The only close to happy experience I have had coding with a small local model has been with Pi. Here is how I ran Pi and I was impressed how well it worked with a 9B model:

OLLAMA_CONTEXT_LENGTH=16384 ollama launch pi --model qwen3.5:9b

To get close to a happy experience with Claude Code or OpenCode with small local models I had to use 28B to 32B local models.

Slop.

Reply on news.ycombinator.com