← Coding

模型与开放条件Model & openness

支持云端与本地 LLM，可连接几乎任何模型，明确包含本地模型选项。

官方原文："README 当前推荐的模型是 Claude 3.7 Sonnet、DeepSeek R1 & Chat V3、OpenAI o1 / o3-mini / GPT-4o"

注意：这份推荐列表停留在上一代模型。

这与该项目 13 个月未发 release 的状态一致。

参考时需自行核对当前可用模型。

Supports cloud and local LLMs — can connect to almost any model, with an explicit local-model option.

Official wording (translated): "the models the README currently recommends are Claude 3.7 Sonnet, DeepSeek R1 & Chat V3, OpenAI o1 / o3-mini / GPT-4o"

Note: this recommendation list is stuck on the previous generation of models.

That is consistent with the project not having shipped a release for 13 months.

Verify currently available models yourself when referring to this list.

运行位置Where it runs

本地 Python 进程，在终端内运行，生命周期绑定终端会话。

退出终端即结束。

它是 PyPI 包 aider-chat。

官方徽章显示 680 万次安装。

另有 watch 常驻形态（见 IDE 形态条目）。

A local Python process running in the terminal; its lifetime is bound to the terminal session.

Quit the terminal and it ends.

It is the PyPI package aider-chat.

Official badges show 6.8 million installs.

There is also a long-running watch mode (see the IDE-form entry).

本地文件Local files

Repo Map 是本工具的核心机制。

它是一份简洁的整个 git 仓库的地图。

官方原文："Aider makes a map of your entire codebase, which helps it work well in larger projects."

map 里包含最重要的类与函数及其类型与调用签名。

每次用户提出改动请求时，repo map 都随请求一起发给模型。

map 包含仓库文件列表 + 每个文件的关键符号。

文档明确列出两个作用：

• 一是帮助它理解正在编辑的代码及其与代码库其他部分的关系
• 二是帮助它写新代码时复用代码库里已有的库、模块与抽象

支持 100+ 编程语言。

文件访问范围：能编辑的是你 /add 进 chat 的文件。

repo map 则扫描整个 git 仓库。

扫描范围受 .gitignore / .aiderignore 约束：--aiderignore 指定忽略文件，默认 git 根目录 .aiderignore；--subtree-only 只考虑当前子树。

仓库须是 git 仓库。

官方原文："无 git 时会先问你是否新建。"

map 的 token 预算由 --map-tokens 控制、刷新频率由 --map-refresh（auto/always/files/manual）控制。

大仓库的实际耗时与 token 开销官方未给数字，本站未做端到端实测。

Repo Map is this tool's core mechanism.

It is a concise map of your entire git repository.

Official wording: "Aider makes a map of your entire codebase, which helps it work well in larger projects."

The map contains the most important classes and functions with their types and call signatures.

Every time the user requests a change, the repo map is sent to the model along with the request.

The map contains the repository file list plus the key symbols of each file.

The docs explicitly list two purposes:

• helping it understand the code being edited and how it relates to the rest of the codebase
• helping it reuse existing libraries, modules and abstractions when writing new code

Supports 100+ programming languages.

File access scope: what it can edit are the files you /add to the chat.

The repo map scans the entire git repository.

Scanning is constrained by .gitignore / .aiderignore: --aiderignore points at the ignore file (default: .aiderignore at the git root); --subtree-only considers only the current subtree.

The repository must be a git repository.

Official wording (translated): "without git it first asks whether to create one."

The map's token budget is controlled by --map-tokens; refresh frequency by --map-refresh (auto/always/files/manual).

For large repos, actual time and token cost: the official gives no numbers, and this site has not measured them.

关机后的任务Tasks after shutdown

不支持终端形态的后台长任务。

另有 watch 模式（见 IDE 形态条目）：常驻监听文件变化并响应编辑器里的 AI 注释。

但那是本地进程常驻，不是云端后台。

No background long-running tasks in the terminal form.

There is a watch mode (see the IDE-form entry): a long-running process watching file changes and responding to AI comments in the editor.

But that is a local process staying resident, not a cloud background.

工具与扩展Tools & extensions

文件编辑 + Git 操作。

README 强调 Git 集成本身就是核心功能。

它会自动用合理的 commit message 提交。

你也可以用熟悉的 git 工具 diff、管理与撤销 AI 改动。

官方文档另有 docs/languages.md（语言支持）与 docs/git.md。

可配置 lint / 测试命令自动运行：

• --lint-cmd / --auto-lint（默认对编辑过的文件跑 lint）
• --test-cmd / --auto-test
• /run 跑 shell 命令并可选加输出进 chat
• /web 抓网页转 markdown
• /paste 贴图

文档出处是 docs/usage/lint-test.html。

无内置 MCP 支持：官方选项表与文档站导航都没有任何 MCP 条目（已查 config/options.html 与 docs 目录）。

故不能接 MCP server。

File editing + Git operations.

The README stresses that Git integration is itself a core feature.

It automatically commits with sensible commit messages.

You can also use familiar git tools to diff, manage and undo AI changes.

The official docs also have docs/languages.md (language support) and docs/git.md.

Configurable lint / test commands run automatically:

• --lint-cmd / --auto-lint (runs lint on edited files by default)
• --test-cmd / --auto-test
• /run runs a shell command and optionally adds the output to the chat
• /web fetches a web page and converts it to markdown
• /paste pastes an image

The docs source is docs/usage/lint-test.html.

No built-in MCP support: neither the official options table nor the docs-site navigation has any MCP entry (checked config/options.html and the docs directory).

So it cannot connect to MCP servers.

上下文与记忆Context & memory

Repo Map 提供全库结构视图，每次请求随附（见 local_files）。

上下文超限用「摘要压缩」而非简单截断。

官方原文："Soft limit on tokens for chat history, after which summarization begins"

这句是官方选项 --max-chat-history-tokens 的描述。

即超过软上限即开始摘要，摘要由 --weak-model 生成。

每个模型的上下文窗口与费用由 --model-metadata-file 提供。

会话内可用 /tokens 查看用量、/drop、/clear、/reset 腾空间。

默认不跨启动恢复（--restore-chat-history 默认 False）。

chat 历史默认写 .aider.chat.history.md。

Repo Map provides a whole-repo structural view, attached to every request (see local_files).

When context overflows it uses "summarization" rather than simple truncation.

Official wording: "Soft limit on tokens for chat history, after which summarization begins"

That line is the official description of the --max-chat-history-tokens option.

That is: past the soft limit summarization begins; summaries are generated by --weak-model.

Each model's context window and cost are supplied by --model-metadata-file.

In-session: /tokens shows usage; /drop, /clear, /reset free space.

No cross-launch restore by default (--restore-chat-history defaults to False).

Chat history is written to .aider.chat.history.md by default.

权限与限制Permissions & limits

默认自动提交（--auto-commits 默认 True）。

这是它的一体化设计而非附加功能。

--dirty-commits（默认 True）会先把已有的未提交改动也提交。

撤销路径是 git 本身，不是工具提供的回滚功能。

官方原文："Use familiar git tools to easily diff, manage and undo AI changes."

会话内用 /undo。

存在确认环节：官方提供 --yes-always（别名 --yes）。

官方原文："Always say yes to every confirmation"

即默认会对某些操作询问，加此开关才全自动。

--dry-run 可不改文件地预演。

无沙箱机制：官方选项表没有任何沙箱项。

除你配置的 lint / 测试命令（--auto-lint / --auto-test）外，它不会自主执行任意 shell 命令（--suggest-shell-commands 只是建议）。

Auto-commits by default (--auto-commits defaults to True).

This is its integrated design, not an add-on feature.

--dirty-commits (default True) first commits existing uncommitted changes as well.

The undo path is git itself, not a rollback feature provided by the tool.

Official wording: "Use familiar git tools to easily diff, manage and undo AI changes."

In-session use /undo.

A confirmation layer exists: the official provides --yes-always (alias --yes).

Official wording: "Always say yes to every confirmation"

That is: by default it asks about certain operations; only with this switch does it go fully automatic.

--dry-run previews without changing files.

No sandbox mechanism: the official options table has no sandbox entry.

Apart from the lint / test commands you configure (--auto-lint / --auto-test), it does not autonomously run arbitrary shell commands (--suggest-shell-commands only suggests).

适合什么任务Best for

适合三类场景。

留在 git 里的改动：希望 AI 改动始终留在 git 版本控制里。

这样的改动可 diff、可撤销。

多模型接入：需要连接本地模型或多家云端模型的用户。

复用已有抽象：需要「写新代码时复用已有抽象」的人。

这也是 Repo Map 的第二个作用。

不适合需要活跃维护与新模型支持的场景（见下方维护状态）。

Fits three scenarios.

Changes kept in git: you want AI changes to always stay under git version control.

Such changes are diffable and undoable.

Multi-model access: users who need local models or multiple cloud providers.

Reusing existing abstractions: people who want "reuse existing abstractions when writing new code".

That is also Repo Map's second purpose.

Not a fit if you need active maintenance or new-model support (see the maintenance status below).

本页由This page is generated from the ai-handbook 数据层生成（CC BY 4.0）。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.