Aider
by Open-source project
Open-source terminal AI pair programmer for real Git codebases
About
Aider is an open-source command-line tool that acts as an AI pair programmer for your local codebase. Instead of working as an editor extension, Aider runs in the terminal and integrates tightly with your git repository. You launch Aider in any Git-tracked project, then interact conversationally—giving natural language commands such as “Add a feature to parse JSON input in app.py.” Aider will edit app.py using GPT-4 (or GPT-3.5), inject the new code, and present a git diff of the changes for review. You can accept, modify, or reject the changes, with all edits tracked in git—ensuring every AI-generated update is versioned and reversible. Aider excels at complex codebase-wide changes: for example, “Rename the User class to Customer across the codebase” triggers multi-file updates, producing diffs for every modified file before you commit. Developers use Aider to automate large refactors, upgrade dependencies (and handle breaking changes), or perform precise edits, all just by describing the task in plain English. Its conversational interface supports iterative development: after an initial edit, you can chain requests (“Now add error handling there”), and Aider will apply further changes and show you updated diffs. Pricing & Requirements: - Free and open-source. Install via pip; no Aider account needed. - Requires API access to an AI language model. By default, works with OpenAI GPT-3.5 or GPT-4 (user supplies API key via env variable or config). - GPT-3.5 is cheaper and faster, good for simple edits; GPT-4 is best for complex coding/refactoring and understanding context, but costs more per token. - Out-of-the-box it’s built for OpenAI, but advanced users can connect to other LLM endpoints if they alter the source. Special Features: - Partial file edits: Specify exactly which files to edit/consider—Aider never sends your whole repo by default, only the minimum needed. - All code changes shown as git diffs; no auto-commit—full developer control over merging. - Multi-turn conversations: In-terminal chat lets you refine tasks step-by-step (“Actually, add error handling...”). - Suitable for power users and those comfortable in *nix shells/Git, but equally valuable for methodical, auditable AI-assisted development in teams. - All AI-powered edits are version-controlled; every run is transparent and auditable. In summary, Aider brings modern LLM coding capabilities into the safety, reviewability, and reproducibility of your standard git workflow—no vendor lock-in, no cloud integration required except your LLM provider.
What you can do with it
- Refactor legacy modules in a large Git repository with AI-generated edits
- Implement new features across multiple related files from natural language instructions
- Fix failing unit or integration tests by analyzing error output and source code
- Bootstrap a new application’s project structure and initial code from a high-level spec
- Generate and revise code snippets while keeping all modifications tracked in Git
Pricing
Unconfirmed
How to access
Installable Python-based CLI (via pip) that you run in your local terminal from within a Git repository; open, no account or waitlist, and integrates with external LLM APIs (Anthropic, OpenAI, DeepSeek, etc.) or local model runtimes like Ollama while working alongside any editor or IDE.
Access is via open-source CLI installed with pip; you run it locally in any Git-tracked project and provide your own API keys for providers like Anthropic, OpenAI, or DeepSeek, or connect to local models via tools such as Ollama. No account, login, or waitlist is required for Aider itself, but you must have credentials with whichever LLM provider you use.
Tips for getting the best results
Install Aider with pip, then run it from the root of a Git-tracked project so it can map your entire codebase and track changes with Git. Configure API keys for your chosen model provider (for example, using command-line flags like --api-key or environment variables) and pick a model well-suited to code such as Claude 3.7 Sonnet, DeepSeek R1, or GPT-4o. Start by adding only the files relevant to a task rather than the entire repository, then describe your goal in clear, high-level terms and let Aider propose and apply edits while you review Git diffs before committing. Iterate using short, focused prompts, and combine Aider with your usual test runner so you can ask it to fix failing tests after seeing error output; for cost control with remote models, avoid sending large, unnecessary files and consider using local models for exploratory work.
Known limitations
Aider’s capabilities and code quality are constrained by whichever external model you choose, so results vary with model strengths and weaknesses. Because it sends code to remote APIs unless you use local models, there can be latency, token limits, and potential confidentiality concerns depending on provider policies. It requires a functioning Git repository and basic comfort with the command line, which may be unfamiliar to some developers. There is no built-in cost cap: when using paid APIs you must monitor your own token usage and spending, and adding too many large files can make sessions expensive or hit context limits. Complex, multi-step refactors still need careful human review, as the AI may introduce subtle bugs, incomplete changes, or style inconsistencies across a large codebase.
Model / Technology
CLI orchestrator over external LLM APIs and optional local models
Commercial use
Aider itself is open source, so its code is typically usable under its repository license terms, but commercial use of generated outputs is governed by the terms of the underlying LLM provider (such as OpenAI, Anthropic, or DeepSeek); users should review each provider’s API terms and data usage policies for any restrictions, attribution requirements, or differences between free and paid tiers.
Training data
Aider does not train its own language models; it relies on external LLM APIs and optional local models, whose training data typically consists of mixtures of licensed data, human-created content, and web-scale corpora as described in each provider’s documentation. Any controversies or restrictions around training data and data retention policies are specific to those providers rather than to Aider itself.