Build a pack

Teach

A pack is a folder

In chapter 5 you installed a pack someone else built. This chapter is the other side: you write one. A tool pack is an ordinary folder with three things in it:

• .toolpack.yaml — the manifest: an id, a version, a one-line description
• __init__.py — a register(registry) function that hands your tools to the agent
• tools.py — the tools themselves, plain Python functions

The docstring and the type hints of each function ARE what the agent reads to decide when to call it, so write them for the model: say what the tool returns and when it is the right one.

Your id is your namespace

Pack ids look like yourname.weather. The part before the dot is yours; names that start with aither or adk belong to the platform and are refused, so a community pack can never shadow one the platform ships.

Four verbs, one loop

adk pack has four author verbs, and each one answers a single question:

| verb | question it answers | |---|---| | new | what does a working pack look like? | | validate | is the folder correct — without running any of its code? | | dev | does it load the way an agent will load it, and which tools does it expose? | | build | what exact bytes am I publishing? |

validate never imports your code, so you can run it safely on a pack you downloaded from a stranger. It also looks for anything shaped like an API key and reports the file and line without printing the value. dev is the one step that runs your code, and it says so.

Reproducible bytes

build writes a .tar.gz and its sha256. The archive has sorted entries and zeroed timestamps, so the same source always produces the same bytes. Anyone who rebuilds your tagged source gets your digest, which is how a stranger can trust a pack they did not write.