Show HN: Pi pod – Run your pi coding agent in sandboxes on your own server | AI Paper Digest

TL;DR Highlight

Open-source tool runs pi coding sessions in isolated sandboxes on your own server

Who Should Read

Backend or platform developers who want to run the pi coding agent in isolation on a team server

Core Mechanics

pi pod runs pi in sandboxes
- pi sessions run in isolated environments (pods) on a server you operate.
- It describes the environments as composable.
Keeps the familiar pi feel
- The goal is that it feels no different from the pi harness you already tuned.
- Only minimal, curated tooling is added.
Four pieces of a minimal org setup
- Agent sandboxing, native clients, RBAC session sharing, and browser/native app automation.
- pi pod is an ongoing effort to consolidate these.
Self-hosting is the default
- It is designed to be easy to run yourself, stressing ownership and privacy.
- It also mentions businesses keeping the agent environment in their own data center.
Hosted service is coming
- Described as coming soon, with an email signup.

Community Reactions

Opinion that containers aren't a safe boundary
- A comment says to use a MicroVM for untrusted code.
- Another asks what 'isolated sandbox' means exactly and how it differs from a VM.
Pushback: systemd-nspawn or LXC is enough
- One comment criticizes reinventing existing tools.
- Others ask how it differs from LXC/VMs in Proxmox or pi's own sandbox.
Experience running pi in Docker
- A user runs pi in a minimal Docker container at home and asked what pi pod adds.
- Another shared bakedpi, combining Forgejo and pi in Docker.
Questions on team standardization and network isolation
- A commenter asked if pi config can be standardized per team or project.
- Another asked how network isolation between pods is handled.

How to Apply

Audit your agent's current boundary
- In 30 minutes, write down which host files and network the coding agent can reach.
- Note whether it runs in a container or a VM.
Spin up the pi pod repo locally
- Clone github.com/pi-pod/pipod and follow the README.
- Check that your existing pi config still works.
Compare with your Docker/LXC setup
- Compare sandboxing, session sharing (RBAC), clients, and browser automation.
- Adopt pi pod only for the gaps.

Try It

copy & run

# Paste into Claude or ChatGPT (the source has no runnable code, so this is a comparison prompt)

Our team runs coding agents like this:
{{current setup: e.g. Docker container on a server}}

Build a checklist on four items: sandbox isolation level, native clients, RBAC session sharing, browser/app automation.
For each, give current state, risk, and needed improvement in a table.

Next: ask when a MicroVM is needed instead of a container.

Terminology

harnessThe runtime wrapper around an LLM that manages tool calls and sessions. Think of it as the chassis and dashboard, not the engine.

Sandbox (pod)An isolated space so a misbehaving agent can't affect the outside. Like a playpen fence.

RBACRole-based access control. Like access cards by job level.

MicroVMA lightweight virtual machine with stronger kernel-level isolation than containers.

Self-hostingRunning it on a server you operate instead of someone else's cloud.

Show HN: Pi pod – Run your pi coding agent in sandboxes on your own server