Earendil ships Pi 1.0 with native MCP support via Codemode, plus experimental Pi Durable

So how are people actually using Pi? I'm sitting here juggling Claude Code and Codex in a terminal like a caveman.

I run it headless for reviews in CI. That's the whole setup.

I built my own off another popular harness and I know plenty of others doing the same. I ended up putting a GUI wrapper on it because the TUI scares off new users.

You're probably more productive than half the Pi users out there. I spent 20 years in vim and then in Claude and Codex TUIs, and I dropped all of it for a GUI multiplexer because I care about my own output.

The batteries-included kit people point to is oh-my-pi on GitHub. Saves you the first week.

It's been great for local models because it's so light. OpenCode never worked as well with stuff running on consumer hardware.

There's a crowd that likes polishing tools more than shipping with them. Same energy as the Hyprland desktop ricers.

You just type pi and go. The only thing I configured is the default model.

Vanilla Pi is great. I wrote a subagent and Ralph loop extension and that's it. I'm about 75% pi, 20% autolith, 5% my own harness.

I have pi on a VPS listening to a WhatsApp group for bills, dumping them into Django, then a second pi with deepseek OCRs the PDFs and fills in amount and due date. The harness matters because it can crop the PDFs with linux tools before parsing.

My favorite part is sandboxing: pi runs on my laptop and keeps transcripts in one searchable place, but all bash and filesystem calls go to a VM. I still wish subagents and web search were built in instead of extensions.

You're not losing much. I've used everything, and it's mostly window dressing plus token usage deltas. It only starts to matter when you're productizing AI for end users.

I barely use Pi for writing code. I use it as a base for agents and much prefer it to an agent SDK. Someone could build a great personal setup on it for the same reason.

Pi is the vim or mechanical keyboard of the pre-AI era. It doesn't matter, and it will generate infinite discourse anyway.

I'll try it when a local model is good enough for coding. Until then I'm staying on opencode.

Local Qwen 3.5 on CPU handled small tasks fine for me.

I don't get the criteria for what counts as proven. Jev took off less than a month ago and it's already in, MCP has been growing for nearly two years and it just got support.

The 07-28 MCP spec is quite different from earlier versions, so the delay makes sense to me.

Armin from Earendil here. Fair question, and the answer is a bit disappointing: we look at what the models are doing. Codex uses responses lite internally and relies on codemode for parallel tool calls, so codemode was a given. The pi-ai SDK already supported image and classifier models, but there was no way to hook them into the coding agent.

Classification models have existed for almost a century. What's new about Jev is that it's a general purpose classifier, so you skip training your own.