Aider — The Terminal AI Coding Agent That Actually Commits

Aider is what happens when someone decides the IDE is optional. It’s a command-line AI coding assistant that opens your terminal, connects to the LLM of your choice, edits files across your whole codebase, and commits every change to git — automatically, with a generated commit message.

No extensions to install. No IDE to open. No subscription to the tool itself. Just a Python package, an API key, and a terminal.

How It Works

You launch Aider in any git repo:

pip install aider-install
aider-install
aider --model claude-sonnet-4-5

Then you talk to it:

> Add rate limiting to the /api/upload endpoint. Max 10 requests per minute per IP.

Aider reads your codebase, figures out which files are relevant, makes the edits, and commits them — all without you leaving the terminal. It’s genuinely different from what IDE plugins do.

Architect Mode

The standout feature. Architect mode uses two models in sequence: a capable “architect” model plans the changes and writes specs, then a cheaper “editor” model executes the actual file edits.

aider --architect --model claude-sonnet-4-5 --editor-model claude-haiku-4-5

You get Sonnet-level reasoning at Haiku editing costs. For large refactors, the cost savings are significant. For complex problems, the two-pass approach also catches more edge cases than a single model doing both jobs.

The Repo-Map

Aider doesn’t just dump your whole codebase into context (that would be prohibitively expensive). Instead, it builds a “repo-map” — a structured summary of every file: classes, functions, signatures, and relationships — and uses that to understand your project before deciding which files to actually read and edit.

The result is that Aider can reason about a 50,000-line codebase without exceeding context limits or costing a fortune. It’s one of the more technically interesting parts of how the tool works.

Git Integration

Every Aider edit is committed automatically. The commit messages are generated by the LLM based on what changed. After a session, your git log looks like deliberate, well-described work — not a single sprawling “wip” commit.

You can also undo the last commit with /undo, which makes iterating on AI suggestions fast and safe.

Model Freedom

Aider supports over 100 LLMs through LiteLLM under the hood. You can use:

• Claude Sonnet 4.5 / Opus 4.6 — best results for complex reasoning
• GPT-5 (Mini/Nano) — solid alternatives with different cost/capability tradeoffs
• Gemini 3.1 Flash — cheap and fast for bulk edits
• Local models via Ollama — air-gapped or cost-free for simpler tasks

Switch models per session with a single flag. No re-configuration, no subscription change.

Who It’s For

Aider is for developers who are comfortable in the terminal and want an AI coding tool that doesn’t require giving up their editor or their workflow. It fits particularly well with:

• Vim/Neovim/Emacs users who aren’t switching to a new IDE
• DevOps and backend engineers doing scripting, config, and infrastructure work
• Solo founders and freelancers who want full control over AI costs
• Anyone who wants to own their stack — open source, no vendor, no subscription

If you want inline autocomplete while you type, look at GitHub Copilot or Cursor. Aider doesn’t do that — it’s a different model of AI assistance entirely.

Benchmark Performance

Aider consistently ranks among the top performers on SWE-bench (real GitHub issue resolution), particularly when configured with Architect mode and Claude. Paul Gauthier publishes benchmark results regularly at aider.chat/docs/leaderboards.

The tool’s focus on real-world coding tasks — not autocomplete — means the SWE-bench metric is actually meaningful for Aider in a way it isn’t for pure autocomplete tools.