Aider: AI pair programming in your terminal

Original Article ↗ Hacker News Discussion ↗

Overall impressions of Aider

  • Many commenters find Aider “amazing” and use it as a primary AI coding tool alongside ChatGPT or Copilot.
  • Especially praised for understanding unfamiliar codebases and answering hard-to-grep questions like “what code processes this URL?”.
  • The CLI workflow (explicitly declaring files in play, auto-applying diffs, auto-committing) is seen as a big improvement over copy‑paste with web UIs.

Context, repo understanding, and workflow

  • Aider builds a “repository map” so the model has high-level codebase context, not just the edited file.
  • It auto‑generates git commits; disabling this is possible but considered slower.
  • Some criticize its commit messages as describing changes rather than intent.
  • For large changes, users are advised to break work into small, guided steps, “like with a junior dev.”

Comparisons to other tools (grep/LSP/IDE/agents)

  • Compared to grep/find, Aider is praised for semantic queries; power users argue grep/ripgrep are still extremely effective when you know what you’re looking for.
  • Some prefer stable, editor-agnostic tools like ctags or dislike complex LSP setups; others report LSP fragility (e.g., Emacs + python-lsp performance issues).
  • Several alternative tools are discussed: Plandex (task/plan‑oriented, server-based, git-style CLI, better for multi‑step feature work than whole‑repo understanding), cursor.sh, Continue, Cody, Supermaven, OpenInterpreter, and various desktop apps/Emacs packages.

Installation, Python, and environment issues

  • Aider’s Python dependency stack can conflict with project requirements; use of pipx or virtualenvs is suggested to isolate it.
  • Some worry about entangling dev tooling with production dependencies; others note this is a general Python packaging problem, not Aider-specific.
  • Nix packaging exists to simplify installation.

Quality, limitations, and skepticism

  • Experiences vary: some report Aider + GPT‑4 successfully building or refactoring across many files; others see frequent subtle errors or broken code, especially in non‑Python languages.
  • Benchmarks show some newer GPT‑4 Turbo variants performing worse and being “lazier.”
  • Several commenters argue LLMs are best for boilerplate, exploration in “unknown territory,” quick one-off tools, or tedious refactors, not complex, correctness‑critical work.
  • Strong skeptics view LLM-based coding as unreliable autocomplete lacking true reasoning, warning against over-trusting it for nontrivial changes.