Vendor: paul-gauthier
Aider is the premier terminal-based AI pair programming agent, pioneering Tree-sitter-based Repository Map prompt engineering, automated atomic git commits, and benchmark-leading multi-file editing precision.
| Repository | paul-gauthier/aider |
|---|---|
| GitHub Stars | ★ 48.5k |
| Forks | 4.9k forks |
| Primary Language | Python |
| License | Apache-2.0 |
| Technical Domain | AGENTS |
$ bash
pip install aider-chat
Aider (paul-gauthier/aider) is the premier open-source terminal AI pair programming agent developed by Paul Gauthier, consistently ranking among top performers on the rigorous SWE-bench software engineering benchmark. Unlike conventional code autocomplete extensions that operate strictly on isolated snippets, Aider acts as a fully capable autonomous development partner capable of comprehending complex multi-file codebases, performing cross-module refactoring, executing test suites, and authoring atomic git commits directly in local terminal environments.
At the core of Aider's prompt engineering architecture is its pioneering 'Repository Map' technology. Using Tree-sitter to parse abstract syntax trees (ASTs) across an entire project, Aider maps out all class definitions, function signatures, and dependency relationships. It then applies PageRank graph analysis relative to the active task context to compress repository topology into a highly concise structural prompt map. This allows frontier LLMs to navigate massive codebases with minimal token consumption while maintaining holistic architectural awareness.
Aider also pioneered robust output formatting strategies tailored to model architectures (such as Unified Diff patches, Wholefile formats, and optimized Search/Replace blocks). Deeply integrated with local Git repositories, Aider automatically stages and commits successful edits with descriptive commit messages. When edits introduce syntax errors or broken tests, Aider automatically captures stdout/stderr, engages an error-reflection loop, and iterates until the code runs cleanly.
Aider was created out of deep frustration with the gap between AI code generation demos and real-world software engineering realities. Most AI coding assistants operate as glorified autocomplete widgets or isolated chat sidebars, failing when confronted with multi-file refactoring where modifying an interface in Module A requires synchronized updates across Modules B and C while maintaining passing test suites.
Paul Gauthier unified the command line, local Git version control, and frontier reasoning models into a cohesive pair programming agent. By stripping away IDE clutter and focusing relentlessly on rigorous prompt engineering—specifically the Repository Map, precise Search/Replace edit formats, and automated commit safety nets—Aider elevated AI from a passive snippet generator into an active collaborator accountable for compiling code and green tests.
Under the hood, Aider models a codebase as a weighted directed graph. Tree-sitter extracts call graphs and symbol declarations, while PageRank calculates the relative architectural importance of every symbol with respect to the currently edited files. By injecting only the top 5% most salient signatures into the prompt, Aider maintains razor-sharp architectural awareness while bypassing context bloat.
In complex multi-file architectural refactorings, developers give high-level instructions (e.g., 'migrate all synchronous HTTP requests to an async httpx client'). Aider locates the relevant files, applies coherent edits, and maintains syntactic consistency across modules.
In test-driven development (TDD) and bug-fixing loops, launching aider --test 'pytest tests/' enables autonomous self-healing: Aider captures failure tracebacks, diagnoses root causes, and edits source code until all tests pass.
In exploring massive legacy codebases, Aider's Repo Map allows engineers to query repository architecture, data flow, and module boundaries, returning accurate answers referencing exact files and line numbers.
In rapid prototyping, developers describe new feature requirements directly in the terminal, and Aider scaffolds files, dependencies, code implementations, and unit test suites interactively.
Install Aider via pip or pipx:
pip install aider-chatNavigate to your Git repository and launch Aider with your preferred model API key (Claude 3.7 Sonnet or GPT-4o recommended):
export ANTHROPIC_API_KEY="your-api-key"
cd my-project/
aiderOnce the interactive session starts, add files and issue commands:
# Add files to the active context
/add src/auth.py tests/test_auth.py
# Issue an engineering prompt
> Implement JWT refresh token rotation in auth.py and add full pytest coverage.
# Run tests and allow Aider to iterate autonomously
/test pytest tests/test_auth.pyTo revert all changes made in the last interaction, type a single command:
/undoIn production practicality, Aider is remarkably safe and non-destructive. Operating strictly within Git repositories, every code modification is isolated into an atomic Git commit with a descriptive message. If a proposed change is unsatisfactory, typing /undo instantly reverts the working tree to the prior commit state.
In context efficiency, the Repository Map is an engineering masterclass. For repositories containing thousands of files, Aider injects a structural AST index using only 1,000–2,000 tokens, reducing context costs by over 95% compared to naive whole-file concatenation while preserving global architectural context.
Regarding model compatibility, Aider tunes its prompt edit formats per model family—employing tailored diff prompts for Claude, whole-file formats for localized models, and specialized reasoning scaffolding for DeepSeek-R1.
With over 48,000 GitHub stars, Aider is one of the most widely adopted developer agent tools in production today, utilized daily by professional software engineers and core maintainers worldwide. On the SWE-bench benchmark, Aider consistently ranks at the top among open-source agentic harnesses.
Numerous engineering teams have reported completing massive repository-wide framework migrations (such as moving enterprise Python stacks from Flask to FastAPI) in hours rather than weeks with flawless test suite passes.
Aider's Repository Map prompting architecture and Search/Replace diff formats have influenced subsequent open-source coding agents and IDE integrations across the industry, establishing the canonical standard for terminal-based pair programming.
Aider - AI Pair Programming in Terminal with Repository Map Prompting is an open-source AI project developed primarily in Python under the Apache-2.0 license. Aider is the premier terminal-based AI pair programming agent, pioneering Tree-sitter-based Repository Map prompt engineering, automated atomic git commits, and benchmark-leading multi-file editing precision.. Aider was created out of deep frustration with the gap between AI code generation demos and real-world software engineering realities. Most AI coding assistants operate as glorified autocomplete widgets or isolated chat sidebars, failing when confronted with multi-file refactoring where modifying an interface in Module A requires synchronized updates across Modules B and C while maintaining passing test suites. Paul Gauthier unified the command line, local Git version control, and frontier reasoning models into a cohesive pair programming agent. By stripping away IDE clutter and focusing relentlessly on rigorous prompt engineering—specifically the Repository Map, precise Search/Replace edit formats, and automated commit safety nets—Aider elevated AI from a passive snippet generator into an active collaborator accountable for compiling code and green tests. Under the hood, Aider models a codebase as a weighted directed graph. Tree-sitter extracts call graphs and symbol declarations, while PageRank calculates the relative architectural importance of every symbol with respect to the currently edited files. By injecting only the top 5% most salient signatures into the prompt, Aider maintains razor-sharp architectural awareness while bypassing context bloat.
Install Aider via pip or pipx:
pip install aider-chatNavigate to your Git repository and launch Aider with your preferred model API key (Claude 3.7 Sonnet or GPT-4o recommended):
export ANTHROPIC_API_KEY="your-api-key"
cd my-project/
aiderOnce the interactive session starts, add files and issue commands:
# Add files to the active context
/add src/auth.py tests/test_auth.py
# Issue an engineering prompt
> Implement JWT refresh token rotation in auth.py and add full pytest coverage.
# Run tests and allow Aider to iterate autonomously
/test pytest tests/test_auth.pyTo revert all changes made in the last interaction, type a single command:
/undoAider - AI Pair Programming in Terminal with Repository Map Prompting is well-suited for Large Codebase Multi-File Refactoring, Automated Bug Triage & Test Fixing, Terminal Natural Language Development, Legacy Codebase Exploration. With an overall rating of 4.9/5, it offers strong community activity, reliable performance, and easy integration with existing AI pipelines.
In production practicality, Aider is remarkably safe and non-destructive. Operating strictly within Git repositories, every code modification is isolated into an atomic Git commit with a descriptive message. If a proposed change is unsatisfactory, typing /undo instantly reverts the working tree to the prior commit state. In context efficiency, the Repository Map is an engineering masterclass. For repositories containing thousands of files, Aider injects a structural AST index using only 1,000–2,000 tokens, reducing context costs by over 95% compared to naive whole-file concatenation while preserving global architectural context. Regarding model compatibility, Aider tunes its prompt edit formats per model family—employing tailored diff prompts for Claude, whole-file formats for localized models, and specialized reasoning scaffolding for DeepSeek-R1.
Minimal tool for running large language models locally