---

name: aider description: >- Aider is an open-source terminal AI pair programmer that edits files in your git repository with an LLM and commits each change. Use when a user asks to run Aider, let an AI refactor or fix code from the command line, script Aider with --message, auto-run lint and tests after edits, or point Aider at Claude, GPT, Gemini, DeepSeek or a local model. license: Apache-2.0 compatibility: "Python 3.10-3.12 (aider-install can fetch 3.12), git, an LLM API key or local model" metadata: author: terminal-skills version: "1.1.0" repository: https://github.com/Aider-AI/aider category: development tags: - ai-coding - terminal - code-generation - pair-programming - git

Aider — AI Pair Programming in Your Terminal

Overview

Aider runs in a git repository, sends the files you add (plus a map of the rest of the repo) to an LLM, applies the returned edits to your files, and makes a git commit for each change. It works with Claude, GPT, Gemini, DeepSeek and local models through LiteLLM. Latest release checked: 0.86.2 (February 2026).

Instructions

Install

python -m pip install aider-install
aider-install            # installs aider in its own environment (adds Python 3.12 if needed)

# or with uv directly
uv tool install --force --python python3.12 --with pip aider-chat@latest

Aider also publishes a curl ... | sh installer; prefer the package-manager routes above. Plain pip install aider-chat into a project environment works but can clash with your dependencies; it supports Python 3.10 to 3.12.

Provide a model and key

export ANTHROPIC_API_KEY=...        # or OPENAI_API_KEY, GEMINI_API_KEY, DEEPSEEK_API_KEY
aider --model sonnet                # built-in aliases: sonnet, opus, haiku, 4o, deepseek, r1, flash, gemini
aider --list-models claude          # search known model names
aider --model ollama_chat/qwen2.5-coder:14b   # local model via Ollama (set OLLAMA_API_BASE)

Keys can also live in a .env file in the repo root (keep it git-ignored).

Interactive use

cd ~/work/billing-api
aider src/invoices.py tests/test_invoices.py     # files to edit, given on the command line
aider --read CONVENTIONS.md src/invoices.py      # read-only context

In-chat commands: /add, /drop, /read-only, /ask (discuss, no edits), /code, /architect (planner model plus editor model), /run <cmd> (output into chat), /test <cmd>, /lint, /diff, /undo (revert the last Aider commit), /tokens, /model, /clear, /reset, /exit. Switch the sticky mode with /chat-mode ask. A good loop is /ask to agree on a plan, then /code go ahead.

Scripted use

aider --message "Add a deleted_at filter to list_invoices and update its test" \
      --yes-always src/invoices.py tests/test_invoices.py

aider --message-file migration-task.md --yes-always src/db/models.py

--message (-m) sends one instruction, applies edits and exits. --yes-always answers every prompt with yes: use it only in a clean branch or throwaway checkout. --dry-run shows edits without writing them; --no-auto-commits leaves changes uncommitted; --commit commits pending changes with a generated message.

Lint and test after each edit

aider --lint-cmd "python: ruff check --fix" --lint-cmd "javascript: npx eslint --fix" \
      --test-cmd "pytest -x -q" --auto-test

Linting of edited files is on by default (--no-auto-lint disables it) and uses built-in linters unless --lint-cmd is set. --auto-test is off by default. Commands must print errors and exit non-zero on failure; Aider then tries to fix them. A formatter that exits non-zero when it rewrites files confuses this: wrap it in a script that runs twice.

Configuration

# .aider.conf.yml  (read from home dir, repo root, then current dir; later files win)
model: sonnet
auto-test: true
test-cmd: pytest -x -q
lint-cmd:
  - "python: ruff check --fix"
read:
  - CONVENTIONS.md
map-tokens: 2048

Every option also exists as an AIDER_* environment variable (for example AIDER_MODEL, AIDER_AUTO_TEST). Only OpenAI and Anthropic keys may go in the YAML file; use .env for the rest.

Python scripting

from aider.coders import Coder
from aider.models import Model

coder = Coder.create(main_model=Model("gpt-4o"), fnames=["src/invoices.py"])
coder.run("add type hints to every function")

This API is not officially supported and can change between releases; prefer the --message command line.

Edit in your IDE

aider --watch-files watches the repo for one-line comments ending in AI! (make this change) or AI? (answer this question) and acts on them.

Examples

Fix a bug and prove it with tests

User: "Pagination in the orders list repeats rows when sorted by created_at. Fix it with Aider and run the tests."

git switch -c fix/orders-pagination
aider --model sonnet --test-cmd "pytest tests/test_orders.py -q" --auto-test \
      --message "list_orders repeats rows across pages when sorting by created_at. Add id as a tiebreaker sort key and add a regression test." \
      app/orders.py tests/test_orders.py

Result: Aider prints the diff, runs pytest, retries if it fails, and creates a commit such as fix: add id tiebreaker to list_orders ordering. Review with git show, undo with /undo or git reset.

Plan with one model, edit with another

User: "I want the strongest model to plan a refactor and a cheap one to type it."

aider --architect --model opus --editor-model sonnet --read CONVENTIONS.md src/payments/

Describe the refactor at the architect> prompt; the architect proposes the change, and it is applied automatically unless you pass --no-auto-accept-architect.

Guidelines

• Aider commits every change and by default adds a Co-authored-by trailer; --no-attribute-co-authored-by removes it, and --attribute-author or --attribute-committer mark the git author or committer name instead. Dirty files are committed first (--no-dirty-commits to stop that).
• Run it on a branch; git diff main is your review step. Never combine --yes-always with credentials or deploy scripts in the repo.
• Add only the files that need edits; the repo map supplies the rest. Use /read-only for schemas and conventions.
• Model aliases move as new models ship; run aider --list-models when a name is rejected.
• Each prompt sends file contents to the model provider; do not add secret files, and use a local model for private code.
• Edit formats (--edit-format whole|diff|udiff|diff-fenced) default per model; override only when edits fail to apply.