Aider Review

Paul Gauthier's tool has been running since 2023 on a design nobody has improved on. It lives in your repository, it edits through diffs you can read, and it makes a git commit for every change it lands, so undoing the agent is a revert rather than an archaeology project. It builds a repository map so the model sees structure without being fed the whole tree, and it works with essentially any model you can reach over an API, from Claude and GPT down to something served locally by Ollama. The polyglot benchmark it publishes has become a reference point well outside the project, partly because it is one of the few coding evaluations labs have not obviously trained against.

It is also, by 2026, a tool whose philosophy costs it users. Aider expects you to add files to the chat yourself. That discipline is why its edits are surgical and its token bills are small, and it is also why anyone arriving from an agent that crawls the repository unprompted finds it fussy. It is interactive by nature, which makes it a poor fit for the fire and forget background jobs that now define the category, and getting started is a config file and an API key rather than a login and a button.

What you get for the friction is a free, Apache licensed tool with roughly forty four thousand GitHub stars, a release every week or two, a large share of its own code written by itself, and no vendor in a position to reprice it under you. For anyone who wants the model kept at arm's length and the diffs kept in view, it is still the right answer.