How to use Aider: AI pair programming in your terminal

Aider is the tool for people who want an AI coding agent that lives in the terminal, not in an editor, and who want to own the model it runs on. It describes itself in four words, “AI pair programming in your terminal”, and the two things that make it distinctive both follow from that: it commits every change it makes to git, so nothing it does is ever unrecoverable, and it connects to almost any model, including a free one running on your own machine. For a vibe coder who watches the meter and keeps their hands near git, that combination is the pitch.

Here is the whole answer up front. Install Aider with two commands, run aider inside a git repository, and it pair-programs with you, committing each edit as it goes so /undo can always walk it back. Point it at Claude or GPT if you want the strongest model, at OpenRouter if you want a free tier, or at Ollama if you want zero per-token cost on your own hardware. Everything below is the how, section by section.

Install it

From Aider’s own install docs, read 15 August 2026. The quick start is two commands:

python -m pip install aider-install
aider-install

aider-install also installs a separate Python 3.12 to run Aider with, so it does not disturb your project’s Python. If you prefer, there are three other supported paths: uv tool install --force --python python3.12 --with pip aider-chat@latest, a direct python -m pip install -U --upgrade-strategy only-if-needed aider-chat on Python 3.9 to 3.12, or pipx. Any of the four gets you the aider command.

Run it in your git repo

This is the habit that defines Aider. Start it inside a git repository, and Aider will git commit all of its changes, each with a sensible message, so you can diff, track, and undo them with the git tools you already use. That is not a safety feature bolted on; it is how the tool works. When the model does something you did not want, you type /undo, which the docs define exactly as “undo the last git commit if it was done by aider,” and you are back where you were.

If you are new to it, run aider in a small repo, ask it to make one change, and watch it commit. Seeing the commit appear is the moment Aider clicks, because it means every step is a checkpoint you can return to.

The commands you will actually use

Aider is driven by in-chat slash commands. From the command reference, the ones that carry most of the work:

• /add puts files into the chat so Aider can edit them; /drop removes them again.
• /ask asks a question about the code without editing it.
• /code requests a change; /architect runs the two-model workflow below.
• /run runs a shell command and can feed the output back into the chat.
• /diff shows what changed since your last message; /undo reverts the last Aider commit.
• /model switches which model you are talking to mid-session; /commit commits edits you made outside the chat.

You will spend most of your time in /add then plain requests, with /undo as your safety net and /run to let Aider see your test output.

Point it at any model

Aider is model-agnostic, and this is where it separates from a tool locked to one lab. Its own docs list the models it works best with, Claude, GPT (o3, o4-mini, GPT-4.1), Gemini 2.5, and DeepSeek, and note that it can connect to almost any LLM besides. You supply a key on the command line with --api-key anthropic=<key> or through an environment variable or .env file, and you are running.

Two paths matter for cost. OpenRouter aggregates many providers behind one key and offers free access to a number of models with daily-usage limits, which the docs call out directly, so it is a common way to try Aider without paying. And a hosted model is only one option, because the next section is the free one.

Run a local or free model

This is the most-searched thing about Aider, and the answer is clean. To run a local model through Ollama: set OLLAMA_API_BASE (it defaults to http://127.0.0.1:11434), pull your model and start the Ollama server, then run aider --model ollama_chat/<model>. That is a coding agent running against a model on your own laptop, with no per-token bill, the same zero-cost setup our Ollama guide walks through from the model side.

More broadly, Aider connects to any OpenAI-compatible endpoint: export OPENAI_API_BASE=<endpoint> and export OPENAI_API_KEY=<key>, then aider --model openai/<model-name>. That covers LM Studio, a self-hosted server, or any provider that speaks the OpenAI API. The trade with a local model is the familiar one, capability for cost and privacy, which our cost study frames in full; a strong 14B or 32B coding model is enough for a great deal of real work.

Architect mode: two models, one job

Aider has a distinctive mode worth knowing. In architect mode, an architect model proposes the changes and a separate editor model translates that proposal into specific file edits. The idea is to let a strong reasoning model plan while a fast, edit-reliable model does the mechanical writing. You enter it with /architect, or start Aider with --architect or --chat-mode architect. It is the one feature here that a single-model IDE agent cannot easily copy, and it pairs naturally with the model-mixing the agent field guide maps across the bring-your-own-model camp.

Configure it, and use it beside your editor

Two practical notes for daily use. Configuration lives in an .aider.conf.yml file in your home directory or your repo root, with AIDER_* environment variables and a .env file as alternatives, so you can set your default model and keys once and forget them.

And although Aider is a terminal tool, it works well beside a graphical editor through watch mode. Run aider --watch-files and it watches your repo for one-line “AI comments”: a comment ending in AI! tells Aider to make a change right there, and AI? asks it a question. So you stay in VS Code, drop an # fix this off-by-one AI! comment, and Aider acts on it, which is the closest thing to an inline experience without leaving the terminal running alongside.

Is it free, and how good is it

Aider itself is free and open source under the Apache-2.0 license. The cost is the model: you pay the underlying API’s token bill unless you point Aider at a free tier like OpenRouter’s or a local model through Ollama, in which case running it costs nothing but your hardware. That is the whole economic story, and it is why Aider shows up so often in “how to code with AI for free” answers.

On capability, Aider publishes its own polyglot leaderboard, which tests models on 225 challenging Exercism exercises across C++, Go, Java, JavaScript, Python, and Rust. When we read it on 15 August 2026 the top entry was gpt-5 at high effort with 88.0% correct, and the useful thing is not that single number, which will move as models ship, but that the tool gives you a dated, reproducible way to pick the model you run it on. Start with aider in a small repo on whichever model you can afford, lean on /undo, and you will know within a session whether the terminal-and-git rhythm is yours.