A CLI tool born in 2023
In May 2023, Paul Gauthier, an engineer with a past at Motorola, publishes Aider, a terminal tool for AI pair programming written in Python. The project is among the first — in the era immediately following ChatGPT — to explore a usage model in which the language model is not an external chatbot, but a collaborator integrated into the editor and version control system of the project. The licence is Apache 2.0.
Aider does not include a graphical interface: the interaction flow happens entirely in the terminal, with text commands and code diffs displayed inline. The choice is deliberately close to the working model of a Unix developer.
Native git integration
Aider’s distinctive feature is direct integration with git. Every change suggested by the model and accepted by the user is translated into an automatic commit, with a message generated by the model itself. This approach turns git history into a detailed log of interactions with the AI and allows easy reversal of unwanted changes via git reset or git revert.
Aider supports working on branches: the user can create a dedicated branch for the pair programming session and merge when the result is satisfactory. The tool automatically manages the diff between versions and uses it as part of the model’s context, avoiding the need to resend the entire file at every iteration.
Repo map with tree-sitter
To work on large codebases, Aider introduces the repo map mechanism: a synthetic representation of the entire repository including only the most relevant symbols (functions, classes, public methods) selected through static analysis with tree-sitter. The repo map is passed to the model as context, giving a global view of the project without saturating the context window.
This approach allows Aider to operate effectively on projects with hundreds of files, a limit that blocks many simpler tools based on direct file concatenation into the prompt.
Supported models and benchmark
Aider supports OpenAI, Anthropic, DeepSeek, Ollama, OpenRouter and other providers through a unified abstraction. The project maintains a self-published “aider code editing leaderboard”, in which different models are evaluated on a controlled refactoring benchmark: applying precise changes to Python files while preserving the correctness of unit tests. The leaderboard has been adopted by the community as an informal reference for comparing models on the specific code editing task.
Link: aider.chat