Aider

Open source terminal AI pair programming tool, Git native workflow, supports multiple large models

Vibe Coding

https://aider.chat/

Visit Website

In-depth Report

Aider is an open source terminal AI pair programming tool, released by Paul Gauthier in July 2023. It currently has 42K+ stars on GitHub, and PyPI has been installed more than 5.7 million times. Its biggest feature is Git's native workflow - every time AI modifies the code, it automatically creates a git commit, making all modifications traceable. Aider supports access to multiple large models such as Claude, GPT, DeepSeek, etc., and can even run local models through Ollama without being bound to any manufacturer. It is suitable for heavy terminal users, teams with strict requirements on Git history, and individual developers with limited budgets.
Aider was created and maintained by independent developer Paul Gauthier and is open source under the Apache 2.0 license. The project has continued to iterate since its release in July 2023. As of April 2026, approximately 15 billion tokens are processed every week. 88% of the new code in the latest version is written by Aider itself - this in itself is the best proof of "eating your own dog food". In the AI programming tool ecosystem, Aider belongs to the CLI Agent type, similar to Claude Code, Goose, and Amazon Q Developer CLI. Its core positioning is "AI pair programming in the terminal." Users describe their needs in natural language in the terminal, and Aider reads the code, generates modifications, directly writes files, and automatically git commits. The entire process does not require opening the IDE or copying and pasting code into the chat window.
Git native workflow This is the most underrated feature of the Aider. Every time AI modifies the code, a git commit will be automatically created, and the commit message clearly states what was changed and why. Every modification step is recorded, and you can use git diff or git log to backtrack at any time. If you are not satisfied, you can use git revert to roll back. During code review, AI modifications are as traceable as manual modifications. The refactoring scenario is particularly easy to use: refactor in steps, automatically commit each step, and roll back directly if the changes fail. This "retrospective AI programming" experience is unique among current AI programming tools. Model freedom Aider is not vendor-locked and can be connected to mainstream models such as Claude, GPT, Gemini, DeepSeek, and Kimi. It also supports running local models through Ollama. As long as it is a model compatible with OpenAI API, it can basically be accessed. This flexibility gives developers the freedom to choose models based on task complexity and budget. code base mapping Aider automatically generates a map of your entire code base to help work better on large projects. It will understand the project structure and file relationships, rather than just looking at the current file. This is especially useful for cross-file modifications and refactoring of old projects. Other useful functions Supports 100+ programming languages: Python, JavaScript, Rust, Ruby, Go, C++, PHP, HTML, CSS, etc. Voice programming: supports voice requests for new features, test cases, or bug fixes Image and web support: Images and web pages can be added to chats to provide visual context Linting & Testing: Automatic linting and testing after every change to fix detected issues IDE integration: Use it in your favorite IDE or editor to request changes by adding comments in your code Practical commands Command /add: Add files to the context Command /undo: Undo the last modification (actually git revert) Command /diff: View the currently modified diff Command /run: Execute the command (such as running a test), and the results are fed back to the AI to continue repairing Command /architect: switch to architecture mode, plan the solution first and then change the code /run works particularly well with test-driven development: run the test → get the error report → say "fix it" → AI automatically fixes it. This loop is much more efficient than manually copying and pasting errors into the chat window.
Aider itself is completely free and open source, but you need to connect to the APIs of various LLM providers when using it, which will incur API call fees. This means that the cost is completely controllable - you can choose a cost-effective model (such as DeepSeek), or you can use a local model to achieve zero cost. Suggestions for model selection in different scenarios: Daily coding + refactoring: Recommend Claude 3.7 Sonnet, with strong coding ability and optimal cost performance Complex architecture design: Recommended Claude Opus 4.6, the best reasoning depth, suitable for difficult problems Quick Q&A + minor modifications: Recommended GPT-4o-mini, fast response and low cost Domestic model preference: Recommend DeepSeek V4, good Chinese understanding, cheap price
positive review Good Git integration: automatic commit, traceable modifications, friendly team collaboration Flexible models: not bound to manufacturers, can freely switch between various models Workflow compatibility: Does not interfere with existing editor habits, you can continue to use Vim and Emacs Terminal friendly: suitable for SSH remote development and scenarios without graphical interface Controllable cost: Optional cost-effective model, controllable token usage negative feedback Agent is not as capable as Claude Code: Claude Code is deeply bound to the Anthropic ecosystem and has stronger project understanding capabilities. No GUI in terminal form: Developers who are not used to the command line may find the threshold high. Learning threshold: You need to be familiar with command line operations, and the configuration is relatively complex. The documentation is not complete enough: some functions need to be explored
media opinion WeChat public account "GHub Open Source Selection" commented: "Aider is an AI pair programming tool that runs in the terminal. It has currently gained 31.6k stars on GitHub. What impresses me most about it is that it is completely open source (Apache 2.0 protocol)." In its review of 10 terminal CLI AI programming tools, the WeChat public account "Claw Notes" pointed out: "Aider is the one with the longest history and the most active community among the open source CLI AI programming tools. It has been polished by a large number of real projects and has high reliability." Competitive product landscape Aider: CLI Agent type, the highlight is open source, multi-model support, suitable for self-hosted teams and end users Claude Code: CLI Agent type, the highlight is project-level understanding and large context, suitable for senior developers and complex tasks GitHub Copilot: IDE completion type, the highlight is high integration, fast response, suitable for daily coding Cursor: IDE project-level type, the highlight is conversational programming, suitable for new project development Cline: VS Code plug-in type. The highlight is that it is free and has strong agent capabilities. It is suitable for VS Code users. Comparative analysis Aider vs Claude Code: Claude Code is stronger but more expensive, Aider is cheaper and open source. Claude Code is deeply bound to the Anthropic ecosystem and has stronger project understanding capabilities; Aider is not bound to manufacturers and can freely switch models. Aider vs Cursor: Cursor is an IDE integrated experience, suitable for front-end development and team unified IDE; Aider is a terminal tool, suitable for SSH remote development and scenarios without graphical interface. Aider vs GitHub Copilot: Copilot is a completion tool suitable for daily coding; Aider is an Agent tool suitable for batch modification, cross-file operations, and automated tasks. Recommendations for combination use Some developers have proposed a combination strategy: use Cursor in the IDE to write code daily, use Aider for batch reconstruction or complex logic cutting, and leave architecture-level tasks to Claude Code. Tools are not mutually exclusive, the key is scene matching.
Configuration complexity The configuration of Aider is relatively complicated, requiring setting API Key, selecting model, and configuring environment variables. For developers who are not familiar with the command line, the threshold for getting started is high. However, the community provides a large number of tutorials, and domestic users can also simplify configuration through the proxy API. Agent capability boundary Aider's Agent capabilities are not as good as Claude Code's, which can understand more complex project contexts and execute longer task chains. If the project is very complex, Claude Code or manual intervention may be required. data privacy Aider needs to send the code to the LLM API, if using the cloud model, the code will leave local. For enterprise environments, an on-premises model (via Ollama) or a self-hosted scenario is recommended.
Suitable for whom Heavy terminal users: accustomed to command line workflow, no need for GUI Developers who don’t want to be bound: want to switch models freely and not be locked in by a single manufacturer Teams that have strict requirements on Git history: every modification needs to be recorded and traceable Individual developers with limited budgets: open source and free, with cost-effective models to choose from SSH remote developer: no graphical interface scenario, terminal tools are the only option Not suitable for whom Command line phobia: not used to terminal operations, need GUI Pursuing the strongest Agent capabilities: The project is very complex and requires Claude Code level understanding. Enterprise compliance requirements are strict: code cannot leave the intranet, and a purely local solution is required (Ollama available) Practical Tips Use .aiderignore to control the context: the syntax is the same as .gitignore, excluding node_modules/, *.min.js, etc. Use /architect first when editing multiple files: let AI come up with a plan first, confirm the idea before taking action, and have a high success rate Cooperate with test-driven development: use /run to run the test, get the error report and let AI fix it, forming a closed loop Environment variable management: Use .env files to store API configurations and read them automatically
Aider is the most "Unix philosophy" AI programming tool - do one thing well and use it in combination with other tools. It’s not fancy, but the three characteristics of Git’s native workflow, model freedom, and terminal friendliness make it occupy a unique position in the AI programming tool ecosystem. For heavy terminal users, developers who don’t want to be bound, and individuals with limited budgets, Aider is an option worth trying.

User Reviews

Evelyn440

—

在服务器上用 Aider 改后端，纯命令行没有图形界面，但胜在能直接跑在 SSH headless 环境里。每次改完自动 commit，消息还写得挺像人，git bisect 找回那次 AI 改坏的接口只花了五分钟，这种可追溯性真的是其他工具比不了的。

David.Brooks_Pro

—

用 Aider 配 DeepSeek V3 跑日常重构，一个月 token 费不到十刀，同等活儿换 Cursor 得二十刀月费，省下的都是实打实的。

Donald.Sanchez_88748

—

回不去了，aider 就是开源免费的 Claude Code。

z6vaudzb

—

architect 模式真香，贵的模型负责出方案、便宜的模型负责改代码，复杂重构的质量明显比单模型一遍过要稳，而且省钱。不过小改动就别开它了，纯属增加开销。

DRogers36995

—

repo-map 是真杀招，用 tree-sitter 把整个仓库压成几百行的图，大模型不用把每个文件都塞进上下文，大仓里也不至于 token 爆表。看当前地图敲 /map 就行，还能用 --map-tokens 调大小。

realNellyAlmonte_x

—

公司代码涉密不能出内网，Aider 接本地 Ollama 跑 Qwen3 Coder 就解决了，同样的配置换到个人项目就是 Claude，一套工作流两个环境通吃。这种数据不出本机的合规场景，SaaS 编辑器根本进不来。

PolygonPioneerMyers

—

新手劝退警告，命令语法、文件管理、edit format 这些弯弯绕绕得先啃几天才顺。我前三天对着终端一脸懵，到第三天才突然开窍，要么爱死要么弃坑，中间派几乎没有。建议先从 /add /diff /undo 三个命令摸起。

kq9rvy5yc

—

吐槽一下，没有行内补全，改个小东西也得走聊天流程，这时候还是 Cursor 顺手。

Alexander_GreenX221

—

aider 不支持 MCP 是真的遗憾，我现在串工具链还得自己写脚本，不像 Codex CLI 那边生态那么顺。重度 Agent 工作流的话这点挺吃亏。

John.Chavez_2023

—

纯命令行党直接狂喜，太对味了。

剑客_19

—

token 这事儿得自己盯着，默认没有消费上限，大仓里一个失控的会话烧起钱来挺吓人。我现在都是开 /tokens 看每轮用量，再用 API 后台的账单告警兜底。

i2p9k55by

—

上周拿 Aider 把一个 50 文件的 Python 2.7 老项目迁到 3.12，不到三小时干完，git 历史清清楚楚，Code Review 时同事一句没挑。这就是它最对的场景：大而明确的任务，改完还得能 review、能回滚。

RichardSchneider

—

单 agent 这点比不过 Claude Code，它不会自己开子代理并行干活，复杂多步任务还是 Claude Code 更强。两者互补着用最舒服：aider 做定点编辑，Claude Code 做架构探索。

Kenneth_Williams168

—

voice 模式没想到挺好用，/voice 一说它就改，手头忙的时候真省事。

Diane.ReyesX

—

44k stars 不是白给的。

Emma.Cooper_2023

—

aider 的 leaderboard 已经成了圈内挑模型的参考，排序按最新一列找免费档里分高的就行。DeepSeek 推理模型、Gemini 2.5 Pro 的免费档、Groq 上的 Llama 3.3 70B 都是能白嫖的好搭子。选错模型改不动文件？换个 edit-format 基本就好了。

SophiaGonzalez

—

CONVENTIONS.md 真该早点用，仓库根目录丢一份，每次会话自动加载，等于把你的代码风格、测试哲学、项目坑都喂给 AI。再配个 .aider.conf.yml 写死模型默认值和 lint 命令，启动就是个懂你仓库规则的搭档。这套组合拳下来效率明显不一样。

Rachel.ThomasZ159

—

装的时候踩了 Python 版本冲突的坑，系统 Python、pyenv、虚拟环境搅一块儿了，aider-install 包装器帮我兜底了，但个别奇怪环境还是会失败。装完跑起来没问题，就是进门那一下不太丝滑。

Amber_Carter_X

—

编辑模式四种，routine 用 diff，小文件用 whole，精细活儿用 udiff，复杂计划上 architect。

AnthonyHart_2023

—

/web 这个命令被低估了，直接把官方文档 URL 丢进去，它抓成干净 Markdown 加进上下文，大模型拿到的就是当前 API 文档而不是训练记忆，省了一大堆过时知识的坑。

Stephen598

—

大文件是真的吃 token，仓库里超过 500 行的文件一旦加进上下文就占一大块预算，得想清楚每次 /add 哪些。我现在习惯只丢跟当前任务相关的文件，上下文越紧越快越便宜，这是用 Aider 的基本功。

z0vjh2ornj

—

免费开源还 Apache2.0，真良心。

薛晨明

—

浏览器界面那版还太糙，命令偶尔不执行，体验不稳，老老实实用终端吧。作者说还在实验，但现阶段别指望它替代命令行。

悠然_18

—

aider 从 2023 年中就在做了，比 Claude Code、Cursor、Codex CLI 都早，今年还在稳定发版。最打动我的是它把无聊的部分做扎实了：repo-map、四种编辑模式、原生 git auto-commit、voice 循环。对预算敏感的团队和个人开发者，2026 年它依然是最稳的 CLI agentic coder。

pOLYGONpIONEERnIELSEN

—

没有订阅费对间歇使用的人来说是真优势，用了才花钱，不是不管用不用每月先扣二十刀。全职重度的话 API 费可能超过 Cursor，所以划不划算看你的使用节奏。

Anna_HowardK

—

像仓库里住了个资深工程师。

Rachel_Johnson_99

—

刚把公司一个老服务从同步 IO 改成异步，跨十几个文件那种，先开 architect 模式出方案再让便宜模型落地，质量比一遍过强太多。架构模式就是让你改之前先看计划，能提前纠偏，复杂改动强烈建议开。

Walter_Baker0077

—

用Aider八个月了，现在完全没法适应Cursor那种GUI工具。不是它更好，而是我这种终端重度用户真的离不开这种工作流。

VBennett007

—

太香了

Charles_Harris_Pro

—

Git集成真的绝了，每次AI改代码都自动提交，commit message还写得挺清楚。改崩了直接git revert，比那些GUI工具靠谱多了。

Aider

In-depth Report

User Reviews

Similar Products