Code Puppy

The deciding trait is how little there is to it. One install from the package index and you have a working agent in a shell, with none of the setup ritual that usually sits between you and the first useful answer. For a Python developer that is thirty seconds, not an afternoon.

You are the wrong buyer if you want the tool to hold your hand through a large refactor, because the surface is thin and the polish goes where the polish went. Pick it as a light daily driver. Pick Aider when the edit loop itself is what you are shopping for.

The gap is the exit. It edits across multiple files, and the row records no git operations at all, which means the tool writes changes and offers nothing that marks where a session began. Undoing a bad run is a manual job with a diff and a memory. This is the class of failure that costs an hour and embarrasses nobody publicly, so it never appears in an issue tracker.

What it does right is scope. It does one job, in one place, and it does not pretend to orchestrate anything.

1. Building on Pydantic AI is the most consequential decision on this row. Tool arguments are validated against declared types by a library maintained elsewhere, which eliminates the family of failures where a model returns almost-correct structure and the agent proceeds anyway. 2. The protocol layer is external too, so tool definitions are not this project's invention.

2. Neither choice is novel, and that is the argument in its favour: the design borrows components whose failure modes are already understood. No evaluation is published, and none is claimed.

806 stars and distribution through a package index that reaches every Python developer on earth. Reach is not capture. There is one maintainer, no entity, no price and no hosted component, so nothing here converts attention into a line item.

Moat: none, and the substitutes are numerous and better funded. Likely acquirer: nobody buys a wrapper this thin. Likely path: it stays useful and small, or the maintainer's interest moves and it stops. Position: a fine dependency for an individual, and never the tool a team's process is built around.

The privacy statement is the part I would quote in a review: no calls beyond the model provider you configure. That is the right commitment and it is a sentence in a README, not a data processing agreement, so it survives exactly as long as the maintainer's intention does.

Zero per seat across sixty engineers, with the provider invoice as the only real cost. There is no SSO, no audit log, no console and nothing that runs in a pipeline. Approved with conditions: developer-installed, never on anything holding regulated data.

MIT, so the fork is available and the code is short enough to read on a train. The model list ends with Ollama, which is the entry I check first, and it means the whole loop runs against weights on my own hardware with no vendor in the path.

Tools arrive over MCP rather than through a plugin system somebody invented here, so the servers I maintain are available without a shim. There is no server side to it, so nothing else can drive this agent, which is a limit I notice more than I mind.