Skip to content

pi-atlas

This site maps the pi-devbox ecosystem: everything we have built around the pi coding agent to make it safe, persistent and manageable on several machines at once.

Out of the box, pi is a coding agent that runs in a terminal. Around it we have added:

  • a container image that ships pi together with a curated toolset, so every machine runs the same environment;
  • extensions and toolkits that change how the agent works, such as delegating tasks to isolated child agents and keeping AWS credentials alive;
  • safety guardrails that make the agent ask before it runs something destructive, and keep secrets out of what the model sees;
  • memory at three levels: within one long session, across sessions, and shared by every machine in the fleet;
  • fleet operations: one place to configure each machine, a mailbox that agents on different machines use to coordinate, and an acceptance routine that every machine runs when it first boots a new release.

The site explains what each part does, why it exists and, in outline, how it is built. The details live in the repository that owns each part; every page links to it.

The big picture

flowchart LR
  src["Source repos"] --> ci["Gitea Actions"]
  ci --> hub["Docker Hub<br/>pi-devbox image"]
  hub --> dev["Each device<br/>compose stack"]
  cfg["docker-compose-repo<br/>per-host config"] --> dev
  dev --> palace["Central palace<br/>memory + mailbox"]
  dev --> llm["Model providers<br/>Bedrock, ERDC"]
  1. Source repositories hold the image definition, our extensions, toolkits and skills. See Repositories.
  2. Gitea Actions builds the image for amd64 and arm64 and publishes it to Docker Hub.
  3. Each device runs the image from a compose stack kept in docker-compose-repo, which also supplies configuration shared by the whole fleet.
  4. Agents on every device read and write the central palace, a shared MemPalace server. It holds long-term memory and the coordination mailbox.
  5. The agent talks to model providers, such as Amazon Bedrock and ERDC inference.

How to read this site

If you want to… Start with
understand how the pieces fit Overview
know what is inside the container The image
see how we changed pi's behaviour The harness
know what stops the agent from doing damage Safety guardrails
understand the memory layers Memory
run or add a machine Fleet operations
know why something is the way it is Decisions
write or fix a page Contributing

How current is this?

Each page names the repositories it was written from and the commit it was read at. When a page and its owning repository disagree, the repository is right. Please report the difference, or fix the page (see Contributing).