To all the haters to not knowing what "pi" means in this context. Is mario zechner's ai agent / coding harness [1][2], similar in spirit albeit much less complex to claude code or openai's codex and others.
I wouldn't call it third tier. If anything alongside opencode, and codex, its one of the "first tier" and the only non-VC-backed (at least before mario joined earendil (idk about earendil raising or not)
anyways. Pi is good. I dont need a telegram client for it, but this precisely show why pi is great, because its really easy to extend pi building plugins or modifying the source (yay, open source)
moreso, it's the only harness that maps close to Neovim philosophy of "everything is a plugin." heck, I took a bite at it and it seems fun to build plugins! especially if it's something as silly as warhammer.pi :) (shameless plug, npm: @bpavuk/warhammer.pi)
that thing makes Pi talk like Adeptus Mechanicus. what's more fitting to a machine than voice that's associated with machines?
I prefer Pi to Claude Code these days, and use it most of the time when I can. The interaction and permissions harness is an order of magnitude better than Claude in my experience, just out of the box. I barely have time to read all the model output, much less fine tune some crazy agent harness tooling that changes underneath me daily and might or might not be documented.
I have the llm twean every little piece of pi which is why I love it. And swapping models. Plus not having scroll jank like I got in opencode.
I even tweaked hitting "enter" when nothing is in the input box, it resume the last request. I love having this control and ease to extend and add anything to the ui
Everything else feels bloated. The same model in pi will consume significantly less context for the same task as there's not piles of unchangeable system prompts being passed along (I assume), and at least compared to OpenCode the quality feels better too. I just wish Anthropic let me use my Claude subscription with pi, it's annoying having to switch between them.
(Not to mention the unmatched level of customisability)
I use my Anthropic sub with Pi. They keep saying they're going to disallow that and then walking back the decision. Or at least I think that's what's happening.
I like the idea of being able to use agents from my phone. Termux is nice, but not very good on small screens.
So I tried Hermes, it had first-class telegram support, but then I never use it for coding and was asking it to delegate to codex / pi. There wasn't much I was gaining from hermes, so I stopped using it.
Then I started using Pi more and liked its extensability. Pi is the first software I have forked and developing to meet my needs. I genuinely love how extensible it is.
I was in the market to see any extensions that help me run Pi, and sure enough there were. One was from Mario himself, the creator of Pi (https://github.com/badlogic/pi-telegram), but it was a bridge instead of being a client itself. I need to start a chat in the TUI and then continue it in telegram. I didn't like that. Most others were same.
I'm in a similar boat, I recently discovered Paseo (https://paseo.sh/) which has mostly surpassed my expectations. It has changed my workflows, pi interaction is via a client rather than inside screen. But...I mostly have not missed that portion.
I have zero affiliation other than being surprised by how well it works. There are minor UX improvements that could be made, but it's a solid...harness-harness ?
Aside: The most useful extra feature I wasn't looking for is that Paseo can inject its own MCP tools to allow managing itself. I now use this to allow pi to manage subagents in "detached" mode, meaning it creates another visible, fully interactive session that I can watch and interrupt as needed.
I see that you also forked Pi itself[1]. Out of curiosity, is there a reason you did this instead of implementing those two features in your fork as Pi extensions?
I've managed to implement most of the things I want in Pi through extensions - with one exception (disabling animations, because there's no extension API or setting for that).
So I guess we (and a lot of other people) have had the same problem, which is managing your agents on the go. I decided to build a plugin[1] for my terminal multiplexer (herdr) to access the sessions via PWA served through a tailnet.
check out tsnet if you are building on tailscale anyways. your app can then be a tailnet node that can read tailnet WhoIs data and have access controlled via tailnet ACL
OTOH all my attempts of driving harnesses via chat have been frustrating because they all fail against the fact that I want to see the code and mobile devices are not really good at that.
Better on tablets, but even there I get more value on having final html artifacts pushed on a server and inspecting them, and dashboards.
Is there some obvious use case I am missing? Or tool for inspecting code?
I don’t run openclaw et al, is that a context where you find this valuable?
"A normal DM is one pi session by default." A single message, or a single message thread? Both are pretty extreme trade-offs to make, honestly. I spent a bunch of time looking at how the various Claws do this, and I settled on an idle timeout causing the next message to start a new session, but I also built /continue, /new, and a /resume that shows recent sessions with a telegram keyboard.
This is awesome, I use telegram as my main chat app with all of my friends and coworkers. Now I need to figure out how to dockerize this and run it on my NAS so it's not reliant on my laptop being online.
Maybe can figure out a workflow to make a branch, deploy it and give me a cloudflare tunnel link to test it, and if approved merge the changes into main branch and deploy to prod. The vibes are off the charts.
Off topic, but apart from pi-tui, is there a recommended TUI library that integrates well with the Pi? I want to have a multi pane TUI experience like lazydocker right inside Pi. Pi-tui is a bit limited.
The customer send me a mail, I program, I send the result back the next day, he sends changes he wants, I program again, send them back next day etc. until he is satisfied.
Who is the programmer? Me or my customer.
Now I hook up an AI to my email account without telling my customer. The AI reads the customer‘s mail, programs, sends the results the next day, customer sends changes, etc.
To all the haters to not knowing what "pi" means in this context. Is mario zechner's ai agent / coding harness [1][2], similar in spirit albeit much less complex to claude code or openai's codex and others.
I wouldn't call it third tier. If anything alongside opencode, and codex, its one of the "first tier" and the only non-VC-backed (at least before mario joined earendil (idk about earendil raising or not)
anyways. Pi is good. I dont need a telegram client for it, but this precisely show why pi is great, because its really easy to extend pi building plugins or modifying the source (yay, open source)
peace
1. https://github.com/earendil-works/pi
2. https://pi.dev/
moreso, it's the only harness that maps close to Neovim philosophy of "everything is a plugin." heck, I took a bite at it and it seems fun to build plugins! especially if it's something as silly as warhammer.pi :) (shameless plug, npm: @bpavuk/warhammer.pi)
that thing makes Pi talk like Adeptus Mechanicus. what's more fitting to a machine than voice that's associated with machines?
Going to try this prompt out within OpenCode, thanks for the share!
I guess it's fair to request a star in exchange: https://tangled.org/bpavuk.neocities.org/warhammer.pi
may Omnissiah bless you and your Machine Spirit
I prefer Pi to Claude Code these days, and use it most of the time when I can. The interaction and permissions harness is an order of magnitude better than Claude in my experience, just out of the box. I barely have time to read all the model output, much less fine tune some crazy agent harness tooling that changes underneath me daily and might or might not be documented.
I have the llm twean every little piece of pi which is why I love it. And swapping models. Plus not having scroll jank like I got in opencode.
I even tweaked hitting "enter" when nothing is in the input box, it resume the last request. I love having this control and ease to extend and add anything to the ui
Everything else feels bloated. The same model in pi will consume significantly less context for the same task as there's not piles of unchangeable system prompts being passed along (I assume), and at least compared to OpenCode the quality feels better too. I just wish Anthropic let me use my Claude subscription with pi, it's annoying having to switch between them.
(Not to mention the unmatched level of customisability)
The default system prompt is so lean:
https://github.com/earendil-works/pi/blob/main/packages/codi...
I use my Anthropic sub with Pi. They keep saying they're going to disallow that and then walking back the decision. Or at least I think that's what's happening.
You mean it's annoying being charged 20x more per token for the privilege of costing them less money? ;)
(In my tests, small tasks cost an order of magnitude less in a minimal harness than in Claude Code and Codex)
Is it really necessary to accuse people of being "haters"?
There was no “context” provided whatsoever, not even in the link.
Its in the first comment. Presumably that slipped down the thread as other comments came in.
https://news.ycombinator.com/item?id=48804802
I like the idea of being able to use agents from my phone. Termux is nice, but not very good on small screens.
So I tried Hermes, it had first-class telegram support, but then I never use it for coding and was asking it to delegate to codex / pi. There wasn't much I was gaining from hermes, so I stopped using it.
Then I started using Pi more and liked its extensability. Pi is the first software I have forked and developing to meet my needs. I genuinely love how extensible it is.
I was in the market to see any extensions that help me run Pi, and sure enough there were. One was from Mario himself, the creator of Pi (https://github.com/badlogic/pi-telegram), but it was a bridge instead of being a client itself. I need to start a chat in the TUI and then continue it in telegram. I didn't like that. Most others were same.
So I built my own!
I'm in a similar boat, I recently discovered Paseo (https://paseo.sh/) which has mostly surpassed my expectations. It has changed my workflows, pi interaction is via a client rather than inside screen. But...I mostly have not missed that portion. I have zero affiliation other than being surprised by how well it works. There are minor UX improvements that could be made, but it's a solid...harness-harness ? Aside: The most useful extra feature I wasn't looking for is that Paseo can inject its own MCP tools to allow managing itself. I now use this to allow pi to manage subagents in "detached" mode, meaning it creates another visible, fully interactive session that I can watch and interrupt as needed.
I see that you also forked Pi itself[1]. Out of curiosity, is there a reason you did this instead of implementing those two features in your fork as Pi extensions?
I've managed to implement most of the things I want in Pi through extensions - with one exception (disabling animations, because there's no extension API or setting for that).
[1] https://github.com/atharva-again/pi
So I guess we (and a lot of other people) have had the same problem, which is managing your agents on the go. I decided to build a plugin[1] for my terminal multiplexer (herdr) to access the sessions via PWA served through a tailnet.
[1] https://github.com/AltanS/collie
check out tsnet if you are building on tailscale anyways. your app can then be a tailnet node that can read tailnet WhoIs data and have access controlled via tailnet ACL
really cool! didn't know this existed, thanks! Going to have to figure out how to make this work with headscale though, looks a bit tedious.
Congrats on releasing this.
OTOH all my attempts of driving harnesses via chat have been frustrating because they all fail against the fact that I want to see the code and mobile devices are not really good at that.
Better on tablets, but even there I get more value on having final html artifacts pushed on a server and inspecting them, and dashboards.
Is there some obvious use case I am missing? Or tool for inspecting code?
I don’t run openclaw et al, is that a context where you find this valuable?
"A normal DM is one pi session by default." A single message, or a single message thread? Both are pretty extreme trade-offs to make, honestly. I spent a bunch of time looking at how the various Claws do this, and I settled on an idle timeout causing the next message to start a new session, but I also built /continue, /new, and a /resume that shows recent sessions with a telegram keyboard.
This is awesome, I use telegram as my main chat app with all of my friends and coworkers. Now I need to figure out how to dockerize this and run it on my NAS so it's not reliant on my laptop being online.
Maybe can figure out a workflow to make a branch, deploy it and give me a cloudflare tunnel link to test it, and if approved merge the changes into main branch and deploy to prod. The vibes are off the charts.
Off topic, but apart from pi-tui, is there a recommended TUI library that integrates well with the Pi? I want to have a multi pane TUI experience like lazydocker right inside Pi. Pi-tui is a bit limited.
No.
Use zelilij
Why not just have multiple instances of pi inside one of the many multiplexers out there? Tmux, Zelliij, herdr, for example.
NB: "Pi" in this context refers to an agent harness of some sort, and does not refer to the Raspberry Pi.
I was confused.
I build a Signal client for Swival: https://github.com/swival/nbclaw
And I use it daily for simple tasks.
Did you built it or did you tell PI to build it?
There'd be no shame either way. Pi is built on the idea that you should ask it to extend itself.
But there is a difference in ordering something or doing it.
If PI extended itself, PI did it.
If I tell you, stay hydrated and get enough sleep, did I improve your live or you?
Would I have without your prompt?
Would I program without a customer‘s order.
Imagine it works like this.
The customer send me a mail, I program, I send the result back the next day, he sends changes he wants, I program again, send them back next day etc. until he is satisfied.
Who is the programmer? Me or my customer.
Now I hook up an AI to my email account without telling my customer. The AI reads the customer‘s mail, programs, sends the results the next day, customer sends changes, etc.
Who is the programmer?
Your question makes more comparative sense with the word “developer”.
You know it's AI slop when the title is some variation of "I built.."
Honestly, there is no need for that sort of comment. Its very normal these days to build things with AI and they don't necessarily have to be slop.