Aider is an AI pair programming tool that runs in your terminal to start new projects or modify existing codebases. It integrates with Git for automatic commits, supports voice-to-code, and works with LLMs like Claude 3.7 Sonnet, DeepSeek R1, OpenAI o3-mini, and o1-preview. Installation is performed via `python -m pip install aider-install` or using `uv`. The tool uses tree-sitter for repository mapping and linting, supports various edit formats including unified diffs and SEARCH/REPLACE blocks, and features an experimental browser UI. It benchmarks performance across models such as GPT-4 Turbo, Claude 3.5 Sonnet, Llama 3.1, and QwQ, achieving state-of-the-art results on SWE Bench Lite and polyglot leaderboards.
Tokens
37.4K
Snippets
117
Records
465
Agent score
50%
What's inside Aider
Git integration overview
Aider integrates tightly with git to provide automatic commits of AI edits, instant rollback via /undo, and reviewable git history. It automatically creates a git repo if launched in a non-git directory, commits each file edit with a descriptive message, and handles dirty files by committing preexisting changes before applying AI edits.
Working with LLM web chat interfaces
Aider can connect to most LLMs via API, but web chat interfaces are useful when: workplace policies restrict usage to proprietary web chat, the web chat LLM has unique context or fine-tuning, API costs are prohibitive, or no API is available. This workflow uses the web chat LLM as a "big brain code architect" while running aider with a smaller, cheaper LLM to apply changes to local files.
File session management in aider
The LLM can only see and edit files that have been added to the chat session. Users add files either via the command line when starting aider or by using the /add command in-chat. If the LLM needs to view specific files, aider prompts the user for permission to add them. Files can also be dropped from the session.
Warn about unsupported reasoning settings
Watch the implementation of a warning system that alerts users when applying reasoning settings to models that don't support them. Covers adding model metadata, confirmation dialogs, refactoring, and comprehensive test coverage.
Infinite output concept
LLM providers limit output via output token limits. Aider bypasses this using models that support 'prefilling' the assistant response. When using a compatible model, Aider displays 'infinite output' in the launch announcement. Models that support prefill can be primed to think they started their response with specific text, continuing generation from that point.
Language support overview
Aider supports most popular coding languages because top LLMs are fluent in mainstream languages and familiar with popular libraries, packages, and frameworks. Most languages work well even without specific repo map or linter support.
Polyglot benchmark overview
The polyglot benchmark is a new code editing benchmark by aider that tests LLMs across 6 programming languages (C++, Go, Java, JavaScript, Python, Rust). It contains 225 of the most difficult Exercism coding problems, designed to be much more challenging than aider's original Python-only benchmark. The benchmark aims to clearly distinguish performance between top coding models, with scores ranging from ~5% to 50% for leading models, leaving headroom for future improvements.
Edit formats overview and CLI configuration
Aider uses various edit formats to let LLMs edit source files. Different models work better with different formats, and Aider defaults to the optimal format for most popular models. To force a specific format, use the --edit-format switch. For architect mode, configure --editor-edit-format. Available formats: whole, diff, diff-fenced, udiff, editor-diff, editor-whole.
Information collected by Aider
Aider collects: (1) Information you provide directly - name, email, message contents, and attachments when contacting them; (2) Device information - IP address, device type, identifiers, browser type/version, OS version; (3) Usage information - pages viewed, referrer URLs, visit dates/times; (4) Analytics information - feature/command usage, error data, LLM usage patterns (associated with random identifier, not personal info); (5) Cookies and tracking technologies from third-party partners.
How Aider uses your data
Aider uses collected information to: provide and improve Services; analyze usage patterns and develop new features; communicate with users about updates and support; generate anonymized/aggregate reports; detect and prevent fraud; comply with legal obligations; and for other purposes disclosed at collection time.
Architect/Editor: Separate code reasoning from editing
Aider supports using two models to complete coding tasks: an Architect model describes how to solve the problem, and an Editor model produces specific code editing instructions. This split approach improves benchmark performance—o1-preview as Architect with DeepSeek or o1-mini as Editor achieves 85% SOTA accuracy. The approach works because reasoning models like o1-preview are strong at problem-solving but struggle with formatted code output; passing their solution to a second LLM (the Editor) produces better code edits.
Repo map concept for providing code context to GPT-4
Aider sends GPT a concise map of your git repository that includes all declared variables and functions with call signatures. This repo map helps GPT understand the overall codebase, decipher code with complex dependencies, and generate new code that respects existing abstractions. The map contains a hierarchical list of all files with their symbols (functions, classes, methods, variables). Callables include their signatures. This approach is more token-efficient than sending entire files, fitting more context into the 8k-token window.
