Pi pod

Canonical version: Pi pod.

Pi pod runs sessions of the Pi coding agent in isolated sandboxes ("pods") on a server you host, and you compose the environments they run in. Its author, edverma2, shared it on Hacker News as a Show HN in October 2026, where it got 121 points and 46 comments.

According to the author, lab tools like Claude Code and Codex lock you into one provider, and the provider-agnostic tools charge per token. Neither lets you fully customize the harness, and both run on someone else's infrastructure. Pi pod gives you Pi's hackable AI Agent Harness on your own VM instead.

Features

• Sessions run in containers inside your own VM
• Configuration is layered, with org-, user- and project-level config for Pi and agents
• The terminal UI renders locally on your machine, so keystrokes don't lag over SSH
• Mobile apps for iOS and Android are native but not released yet; you can build them from the Swift and Kotlin sources on GitHub

What the HN discussion raised

Several commenters said containers aren't a security boundary, and pushed for microVMs (Firecracker, Kata, smolvm) or gVisor / systemd-nspawn instead. The author's answer on the isolation model was thin.

Other points from the discussion:

• Build vs. buy. Building your own "good enough" agent sandbox is cheap now, so some would rather roll their own than trust a shared tool. Others see maintaining a harness as a distraction
• The space is crowded (kagent, VibePod, omnigent, openshell, T3 Code, bakedpi...), and everyone seems to be building their own harness or sandbox
• A "cloud computer for agents" is "just a VM on someone else's computer", which is exactly the argument for self-hosting

References

• https://pipod.dev/
• Show HN discussion: https://news.ycombinator.com/item?id=49937304

Related

• Pi Mono
• AI Agent Harness
• AI Agents
• Self-hosting
• Firecracker
• gVisor
• Claude Code

About Sébastien

Ready to get to the next level?