← Coding

ClineCline

Cline Bot

编程Coding 终端Terminal 本地Local 开源Open source partial lifecycle: active 核验Checked 2026-10-23

提供 CLI、桌面和编辑器入口的编码工具。CLI、Desktop App、VS Code 扩展共享同一开源核心。模型也是自选的。

A coding tool with CLI, desktop and editor entry points — CLI, desktop app and VS Code extension share one open-source core, and the model is selectable.

适合与不适合When to use it, when not to

需要同一工具覆盖编辑器、终端、桌面三种工作形态的用户。需要模型自选（不绑定单一厂商）的人。需要 CI/CD 集成（CLI headless）的人。需要「改动可审查、可回滚」保证的人。每次编辑以 diff 呈现，checkpoint 可整体撤销。需要周期性自动化的人（cron 调度，重启后仍生效）。

不适合要求「审批默认最严」的场景—— CLI / headless 默认 --auto-approve true 全程放行，要收紧得显式传 --auto-approve false（或用 PreToolUse hook 拦命令）。

Users who need one tool across editor, terminal and desktop. People who need model choice (not locked to one vendor). People who need CI/CD integration (CLI headless). People who need reviewable, revertible changes — every edit appears as a diff plus a checkpoint to undo wholesale. People who need periodic automation (cron scheduling that survives restarts).

Not a fit for requiring approval to behave identically across all forms — the extension form confirms human-in-the-loop, but behaviour under CLI headless was not verified.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin: 国外出品foreign-made　国外主体（网信办公示名单未见对应备案主体）。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access: 大陆可直连direct OK　官网仓库在 GitHub，大陆直连波动（同日实测 200/000 混合）；npm / PyPI 安装源实测稳定可达。Hosted on GitHub — direct reachability from mainland fluctuates (200/000 mixed on the same day); npm / PyPI install routes measured stable.
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-20 检索 13 批附件，https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: provider 由你自选。官方给出 30+ 家清单，以自带 API key（BYOK）为主。

官方 Model Selection Guide 的 provider 表列出：

Cline —— 免 API key，登录即用
OpenRouter / Anthropic / OpenAI
OpenAI Codex —— 直接用 ChatGPT 订阅，无需 API key
Google Gemini / DeepSeek / Alibaba Qwen / Moonshot / Cerebras
AWS Bedrock
Ollama —— 本地模型

官方另有「all 30+ supported providers」页，还含 xAI Grok、Mistral、Groq、 Together、Hugging Face、Baseten、SambaNova、Nebius 等。

两种接入方式：

CLI 命令 —— cline auth --provider anthropic --apikey sk-... --modelid ...
官方 SDK 支持 OpenAI-compatible 端点 —— providerId: "openai-compatible" + baseUrl + apiKey

⚠ 换 provider 之后的 tool calling 与 structured output 质量，官方没给对比。
You pick the provider. The official lists 30+, primarily BYOK (bring your own API key).

The official Model Selection Guide's provider table lists:

Cline — no API key needed, just log in
OpenRouter / Anthropic / OpenAI
OpenAI Codex — use your ChatGPT subscription directly, no API key
Google Gemini / DeepSeek / Alibaba Qwen / Moonshot / Cerebras
AWS Bedrock
Ollama — local models

The official also has an "all 30+ supported providers" page, additionally covering xAI Grok, Mistral, Groq, Together, Hugging Face, Baseten, SambaNova, Nebius etc.

Two hookup paths:

CLI command — cline auth --provider anthropic --apikey sk-... --modelid ...
the official SDK supports OpenAI-compatible endpoints — providerId: "openai-compatible" + baseUrl + apiKey

⚠ After switching providers, the quality of tool calling and structured output: the official gives no comparison.
运行位置Where it runs: 这是本站里唯一「三种形态全覆盖」的对象。

形态怎么用
CLI npm i -g cline；支持交互式与完全 headless 两种模式，headless 用于 CI/CD 与脚本；可用 --workspace /path/to/repo 指定目录
Desktop App macOS 与 Windows 原生应用；可在任意文件夹运行 agent session；支持定时任务
VS Code 扩展 另有 JetBrains 插件

.clinerules 会被 CLI / VS Code 扩展 / JetBrains 插件三方共同读取。三套形态共享同一开源核心。
This is the only subject on this site covering all three forms.

Form How it runs
CLI npm i -g cline; interactive and fully headless modes, headless for CI/CD and scripts; --workspace /path/to/repo sets the directory
Desktop App native macOS and Windows apps; run agent sessions in any folder; scheduled tasks supported
VS Code extension plus a JetBrains plugin

.clinerules is read jointly by the CLI / VS Code extension / JetBrains plugin. The three forms share the same open-source core.
本地文件Local files: 两个特点：跨项目协调改动，以及可回滚的检查点。

README 的「Edits Code Across Your Project」说的是：读取项目结构、理解文件间关系，然后在代码库里做协调改动。

改动过程中它会盯着 linter 与编译器的报错，主动修掉缺失的 import、类型不匹配、语法错误 —— 那些「用户看不到就会出问题」的东西。

在 VS Code 与 JetBrains 里，每次编辑都以 diff 形式呈现，可审查、可修改、可回滚
所有改动都有 checkpoint 记录，可以整体撤销
官方 Checkpoints 页说明它用 shadow Git repository 跟踪改动， 与你的主 Git 工作流互不干扰，每次工具调用后建一个检查点

读取范围分两档：默认「Read project files」（工作区内）；「Read all files」（工作区外）需要先打开基础开关才生效。项目级忽略用 .clineignore 控制（官方还给了用 hook 强制的做法）。

⚠ 它没有「全量语义索引」这个环节。 内置 search 工具是「Ripgrep-powered codebase search」；官方对大仓库的建议是「用搜索代替大量读文件」，并把任务拆小。（已查 Tools 页与 Context Window Guide。）

官方对大仓库的建议：Use search instead of reading many files
Two features: coordinated changes across the project, and rollbackable checkpoints.

The README's "Edits Code Across Your Project" means: read the project structure, understand relationships between files, then make coordinated changes across the codebase.

While editing it watches linter and compiler errors, proactively fixing missing imports, type mismatches and syntax errors — the things that "go wrong without the user ever seeing them".

In VS Code and JetBrains, every edit is presented as a diff — reviewable, modifiable, rollbackable
all changes carry checkpoint records and can be undone wholesale
the official Checkpoints page explains it tracks changes with a shadow Git repository, not interfering with your main Git workflow, creating a checkpoint after each tool call

Read scope has two tiers: by default "Read project files" (inside the workspace); "Read all files" (outside the workspace) only takes effect after enabling a base toggle first. Project-level ignores are controlled by .clineignore (the official also shows a hook-enforced approach).

⚠ It has no "full semantic index" stage. The built-in search tool is a "Ripgrep-powered codebase search"; the official advice for large repos is "use search instead of reading many files", and to break tasks down smaller. (checked the Tools page and the Context Window Guide.)

The official large-repo advice: Use search instead of reading many files
关机后的任务Tasks after shutdown: 长进程与定时任务都支持。

命令在终端直接执行，输出实时可见 —— 装包、跑构建、跑测试、部署、管数据库。对 dev server 这类长进程，Cline 会在后台继续工作并对新输出做出反应：编译错误、测试失败、服务崩溃都在发生时被捕捉。

Desktop App 形态另有定时任务（schedule routines）：

可用 cron 调度做周期性自动化
例子：每日 PR 摘要、每周依赖检查、代码库健康报告
schedule 在重启后仍保留，且独立于任何终端会话运行

CLI 的 headless 模式则面向 CI/CD。
Long-running processes and scheduled tasks are both supported.

Commands execute directly in the terminal with output visible in real time — installing packages, running builds, running tests, deploying, managing databases. For long-running processes like dev servers, Cline keeps working in the background and reacts to new output: compile errors, test failures and service crashes are caught as they happen.

The Desktop App form additionally has scheduled tasks (schedule routines):

cron scheduling for periodic automation
examples: daily PR summaries, weekly dependency checks, codebase health reports
schedules persist across restarts and run independently of any terminal session

The CLI's headless mode targets CI/CD.
工具与扩展Tools & extensions: 内置工具清单已核到（官方 Tools 参考，ClineCore 运行时）：

bash —— 执行命令
editor —— 查看/编辑文件
read_files —— 批量读
apply_patch —— 打 unified diff
search —— ripgrep 代码搜索
fetch_web —— HTTP + HTML 转 markdown
ask_question —— 问用户

MCP 是它的核心扩展面。 官方有 MCP Marketplace，配置文件是 .cline/mcp.json：

工具经 use_mcp_tool 调用，资源经 access_mcp_resource 访问
传输支持 stdio 与 HTTP
可在 .clinerules 里用「MCP Rules」按触发词分组调度

其它扩展机制：Hooks（PreToolUse / PostToolUse）· Skills · Workflows · plugins（可注册自定义工具，当前限 SDK / CLI / Kanban）。

官方对 MCP Marketplace 的说法：one-click installation experience for hundreds of MCP servers
The built-in tool list has been verified (official Tools reference, ClineCore runtime):

bash — execute commands
editor — view/edit files
read_files — batch reads
apply_patch — unified-diff patching
search — ripgrep code search
fetch_web — HTTP + HTML-to-markdown
ask_question — ask the user

MCP is its core extension surface. The official has an MCP Marketplace; the config file is .cline/mcp.json:

tools are called via use_mcp_tool, resources accessed via access_mcp_resource
transports support stdio and HTTP
"MCP Rules" in .clinerules can group dispatch by trigger words

Other extension mechanisms: Hooks (PreToolUse / PostToolUse) · Skills · Workflows · plugins (custom tools registerable, currently limited to SDK / CLI / Kanban).

The official statement about the MCP Marketplace: one-click installation experience for hundreds of MCP servers
上下文与记忆Context & memory: 项目级规则文件是 .clinerules。

它用来定义项目专属规则：编码标准、架构约定、部署流程、测试要求。 规则会被 CLI、VS Code 扩展与 JetBrains 插件自动读取。 另可用 skills 让模型按需加载特定规则。

Desktop 形态提到「run agent sessions in any folder」，也就是有跨文件夹的会话管理。

上下文机制已核到：接近上限时 Auto Compact 自动摘要。窗口大小由所选模型决定 —— 官方列的 Claude Sonnet 4.5 为 1M、 GPT-5 为 400K、Gemini 2.5 Pro 为 1M+。

跨会话记忆靠 Memory Bank：一套结构化 markdown 文件（projectbrief.md / productContext.md / activeContext.md / progress.md 等），放进仓库、以 .clinerules 指令驱动读写。

官方另有 Focus Chain（todo 注入）与 /newtask、/smol 做上下文交接。

官方对 Auto Compact 的原文：Creates a comprehensive summary... Replaces the conversation history with the summary
The project-level rules file is .clinerules.

It defines project-specific rules: coding standards, architecture conventions, deployment flows, testing requirements. Rules are read automatically by the CLI, VS Code extension and JetBrains plugin. Skills can additionally let the model load specific rules on demand.

The Desktop form mentions "run agent sessions in any folder", i.e. cross-folder session management.

The context mechanism has been verified: near the limit, Auto Compact summarizes automatically. Window size is decided by the chosen model — the official lists Claude Sonnet 4.5 at 1M, GPT-5 at 400K, Gemini 2.5 Pro at 1M+.

Cross-session memory runs on the Memory Bank: a set of structured markdown files (projectbrief.md / productContext.md / activeContext.md / progress.md etc.) kept in the repo, with reads/writes driven by .clinerules instructions.

The official also has Focus Chain (todo injection) and /newtask, /smol for context handoff.

The official wording on Auto Compact: Creates a comprehensive summary... Replaces the conversation history with the summary
权限与限制Permissions & limits: 两级审批设计，是本站里最细的一套。

第一级是 Plan / Act 双模式。 Plan 模式下 Cline 探索代码库、提出澄清问题、列出策略；对齐之后切到 Act 模式执行计划。

第二级是「每次文件编辑与终端命令都需要批准」，也可以切到 auto-approve 让它自主运行。VS Code 扩展形态的表述是「human-in-the-loop approval」。

⚠ 但 CLI / headless 的默认行为是相反的。这点必须知道 （已核官方 CLI Overview 与 CLI Reference）：

--auto-approve <boolean> 默认 true —— 即默认放行工具
--auto-approve false 时，每个工具调用前都需批准
--yolo 跳过全部审批
ACP（编辑器集成）模式默认相反，为 false

另：CLI 没有 shell 命令的允许/拒绝清单。官方建议用 PreToolUse hook 拦截。

CLI 提示语原文：Default to start in act mode with auto-approve enabled VS Code 扩展表述：use tools with human-in-the-loop approval
A two-level approval design — the finest on this site.

Level one is the Plan / Act dual mode. In Plan mode Cline explores the codebase, asks clarifying questions, and lays out a strategy; once aligned, switch to Act mode to execute the plan.

Level two is "every file edit and terminal command needs approval", switchable to auto-approve for autonomous running. The VS Code extension form's wording is "human-in-the-loop approval".

⚠ But the CLI / headless default is the opposite. This must be known (verified against the official CLI Overview and CLI Reference):

--auto-approve <boolean> defaults to true — i.e. tools are allowed by default
with --auto-approve false, every tool call needs approval first
--yolo skips all approvals
ACP (editor-integration) mode defaults to the opposite, false

Also: the CLI has no shell-command allow/deny list. The official suggests intercepting with a PreToolUse hook.

CLI prompt original wording: Default to start in act mode with auto-approve enabled VS Code extension wording: use tools with human-in-the-loop approval
适合什么任务Best for: 适合四类人：

需要同一工具覆盖编辑器、终端、桌面三种工作形态
需要模型自选（不想绑定单一厂商）
需要 CI/CD 集成（CLI headless）
需要「改动可审查、可回滚」的保证 —— diff 呈现 + checkpoint 整体撤销
Fits four kinds of people:

those needing one tool to cover editor, terminal and desktop work forms
those needing model choice (not bound to a single vendor)
those needing CI/CD integration (CLI headless)
those needing a "changes are reviewable and rollbackable" guarantee — diff presentation + wholesale checkpoint undo

头号误解Top misconception

以为是单一形态的 VS Code 插件，实际有 CLI / Desktop / 扩展三套Assuming it is a single-form VS Code extension — it actually has CLI, Desktop and extension
忽略 headless 模式，CLI 形态本身就支持 CI/CD 集成Overlooking headless mode — the CLI form supports CI/CD integration in its own right
以为审批机制一致，扩展形态是 human-in-the-loop，CLI 默认却是 auto-approve 全程放行Assuming approval behaves identically across forms: the extension explicitly has human-in-the-loop, but behaviour under CLI headless was not verified
以为自动提交是设计意图，实际默认每次编辑都要批准，可开 auto-approve 自主运行Assuming auto-commit is by design — by default every edit needs approval, and auto-approve enables unattended runs

未知项清单Known unknowns

换 provider 后的 tool calling / structured output 质量对齐（官方未给对比）
Desktop 形态 0.0.x 的稳定性预期
大仓库在「无全量索引 + ripgrep 搜索」路线下的实际检索质量 > 原「provider 清单 / 工具清单 / 上下文窗口 / 跨会话记忆 / CLI headless 审批」五项已由官方 > docs.cline.bot（Model Selection Guide、Tools 参考、Auto Compact、Memory Bank、 > CLI Overview / Reference）核到，见上方 axes。

Tool calling / structured output quality alignment after switching providers (the official gives no comparison)
Stability expectations for the Desktop form at 0.0.x
Actual retrieval quality on large repos under the "no full index + ripgrep search" route > The original five items — "provider roster / tool list / context window / cross-session memory / CLI headless approval" — have since been verified against the official docs.cline.bot (Model Selection Guide, Tools reference, Auto Compact, Memory Bank, CLI Overview / Reference); see the axes above.

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-08。

类型Type	名称Name	链接Link
repo	Cline · 仓库 README（三形态说明）Cline · repository README (three-form description)	https://github.com/cline/cline
changelog	Cline · Releases（v4.1.21 / desktop-v0.0.37）Cline · Releases (v4.1.21 / desktop-v0.0.37)	https://github.com/cline/cline/releases
docs	Cline · 官方文档Cline · official docs	https://docs.cline.bot
docs	Cline · CLI 子项目 READMECline · CLI subproject README	https://github.com/cline/cline/blob/main/apps/cli/README.md
docs	Cline · Model Selection Guide（30+ provider 清单、BYOK / Ollama 本地模型）Cline · Model Selection Guide (30+ provider roster, BYOK / Ollama local models)	https://docs.cline.bot/core-features/model-selection-guide
docs	Cline · Auto Approve & YOLO（分类审批、工作区内外、命令 safe / 需批准）Cline · Auto Approve & YOLO (category approvals, inside/outside workspace, command safe / needs approval)	https://docs.cline.bot/features/auto-approve
docs	Cline · Checkpoints（shadow Git 仓库、按步快照、三类恢复）Cline · Checkpoints (shadow Git repository, per-step snapshots, three restore classes)	https://docs.cline.bot/core-workflows/checkpoints
docs	Cline · Auto Compact（接近上限自动摘要，替代旧截断）Cline · Auto Compact (auto summary near the limit, replacing the old truncation)	https://docs.cline.bot/features/auto-compact
docs	Cline · Memory Bank（跨会话结构化 markdown 记忆文件）Cline · Memory Bank (cross-session structured markdown memory files)	https://docs.cline.bot/best-practices/memory-bank
docs	Cline · Tools 参考（内置工具清单与工具名）Cline · Tools reference (built-in tool list and tool names)	https://docs.cline.bot/tools-reference/all-cline-tools
docs	Cline · MCP Marketplace（一键安装数百 MCP server）Cline · MCP Marketplace (one-click install for hundreds of MCP servers)	https://docs.cline.bot/mcp/mcp-marketplace
docs	Cline · CLI Overview（headless 触发、--auto-approve 默认 true）Cline · CLI Overview (headless triggering, --auto-approve defaults to true)	https://docs.cline.bot/usage/cli-overview
docs	Cline · CLI Reference（--auto-approve 默认 true、ACP 模式默认 false）Cline · CLI Reference (--auto-approve defaults to true, ACP mode defaults to false)	https://docs.cline.bot/cli/cli-reference

实测记录Measured runs

本档未跑端到端编码任务；能力结论来自官方文档，编码向测量协议见 protocols/coding/_protocol.md。

实测建议：这个工具的实测重点应是跨形态的一致性—— 同一个任务在三套形态下表现是否一致，审批行为是否一致。

This site has not completed measured runs. The test protocol is at protocols/coding/_protocol.md.

Suggested measurement: the measured focus for this tool should be cross-form consistency — whether the same task behaves the same across the three forms, and whether approval behavior is consistent.

三种形态（README 明确列出）Three forms (explicitly listed in the README)

形态	关键能力
CLI	`npm i -g cline`；交互式对话或完全 headless，用于 CI/CD 与脚本
Desktop App	macOS 与 Windows 原生应用；在任意文件夹运行 agent session；定时任务；管理 models / plugins / MCP servers
VS Code 扩展	创建文件、执行命令、浏览网页、使用工具，human-in-the-loop approval

这个三形态覆盖在本分区里是独一份—— 其他对象要么是 IDE + 扩展（Copilot），要么是编辑器单一形态（Zed）。Aider 则是 CLI 为主。

对用户意味着：可以在编辑器里用、可以在终端里跑、可以在桌面端定时跑，同一套核心，同一份配置习惯。

Form	Key capabilities
CLI	`npm i -g cline`; interactive chat or fully headless, for CI/CD and scripts
Desktop App	native macOS and Windows apps; run agent sessions in any folder; scheduled tasks; manage models / plugins / MCP servers
VS Code extension	create files, execute commands, browse the web, use tools, human-in-the-loop approval

This three-form coverage is unique in this category — other subjects are either IDE + extension (Copilot) or a single editor form (Zed). Aider is CLI-led.

What it means for users: usable in the editor, runnable in the terminal, schedulable on the desktop, one and the same core, one and the same configuration habits.

变更记录：两条 release 线Change log: two release lines

版本	日期
`desktop-v0.0.37`	2026-09-26
`desktop-v0.0.36`	2026-09-25
`v4.1.21`	2026-09-24

注意版本号是分开的两条线：

v4.x 是主版本（扩展与 CLI）
desktop-v0.0.x 是桌面端，还在 0.0.x

这意味着桌面端仍早期，稳定性预期要与主版本分开。

Version	Date
`desktop-v0.0.37`	2026-09-26
`desktop-v0.0.36`	2026-09-25
`v4.1.21`	2026-09-24

Note the version numbers are two separate lines:

v4.x is the main line (extension and CLI)
desktop-v0.0.x is the desktop app, still at 0.0.x

That means the desktop app is still early-stage; stability expectations should be kept separate from the main line.

权限设计：明确的人工审批Permission design: explicit human approval

VS Code 扩展形态的 README 原文：

use tools with human-in-the-loop approval

这是本分区里少见的明确表述—— 多数工具不会在 README 首屏就说明工具执行需要人工批准。

CLI / headless 的审批口径已核到（官方 CLI Overview 与 CLI Reference）：提示语写 "Default to start in act mode with auto-approve enabled"， --auto-approve <boolean> 默认 true——headless 运行默认放行工具， --auto-approve false 才改为每次工具调用前需批准。--yolo 跳过全部审批。 ACP（编辑器集成）模式默认相反，为 false。另：CLI 无 shell 命令允许/拒绝清单。官方建议用 PreToolUse hook 拦截。

所以「三形态审批一致」并不成立：扩展形态默认要人工批准， CLI 默认自主放行——跨形态使用时必须按 --auto-approve 显式对齐。

VS Code extension form, README original wording:

use tools with human-in-the-loop approval

This is a rare explicit statement in this category — most tools do not state on the README's first screen that tool execution needs human approval.

The CLI / headless approval stance has been verified (official CLI Overview and CLI Reference): the prompt says "Default to start in act mode with auto-approve enabled", --auto-approve <boolean> defaults to true — headless runs allow tools by default, --auto-approve false switches to approval before every tool call. --yolo skips all approvals. ACP (editor-integration) mode defaults to the opposite, false. Also: the CLI has no shell-command allow/deny list. The official suggests intercepting with a PreToolUse hook.

So "approvals are consistent across the three forms" does not hold: the extension form requires human approval by default, the CLI defaults to autonomous allowance — when crossing forms you must align explicitly via --auto-approve.

值得单独记的两处设计Two designs worth recording separately

① Plan / Act 双模式

Plan 模式下它只探索代码库、提问、列策略，不执行；对齐后切到 Act 才动手。这是「先想清楚再改」的机制化实现。

② 监控 linter 与编译器错误

README 原文：

It monitors linter and compiler errors as it works, fixing issues like missing imports, type mismatches, and syntax errors before you even see them.

它会主动修「类型不匹配」「缺失 import」这类问题 —— 这恰好是本仓库实测任务集里最难判定的部分，值得单独验证。

① Plan / Act dual mode

In Plan mode it only explores the codebase, asks questions and lays out a strategy — it does not execute; it only acts after switching to Act once aligned. This is the mechanized implementation of "think it through before changing".

② Monitoring linter and compiler errors

README original wording: