← Coding

TraeTrae

字节跳动ByteDance

国产China-origin 免费+付费Free + paid 桌面Desktop WebWeb 移动Mobile SOLOSOLO 自定义模型Custom models MCPMCP partial lifecycle: active 核验Checked 2026-10-30

Trae 是字节跳动的三端 AI IDE（桌面/Web/移动）：Desktop 双窗口（Agent/IDE）+ SOLO 模式整段委托，内置多模型可加自定义 key，Privacy mode 与沙箱执行是官方显式给的开关。

Trae is ByteDance's three-client AI IDE (desktop / web / mobile): dual windows on desktop (Agent/IDE) with SOLO mode for whole-task delegation, built-in multi-model plus custom keys, Privacy mode and sandboxed execution as official switches.

适合与不适合When to use it, when not to

适合：三端同一任务流、SOLO 整段委托、多模型免配 key 起步。不适合：假设国内外两版功能一致、要完全离线。

Fits: one task flow across desktop+web+mobile, whole-task delegation via SOLO mode, starting with built-in multi-model without configuring your own keys. Not a fit: expecting the domestic and international editions to match (two separate lines); fully offline use (cloud tasks are part of the product semantics).

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 2026-10-21

出品属地Origin: 国产domestic　字节跳动（北京字节跳动科技有限公司体系）；国际版与国内版两条线。ByteDance; two lines — international (trae.ai) and domestic (trae.cn).
国内使用Direct access: 大陆可直连direct OK　trae.ai / www.trae.ai / trae.cn / docs.trae.ai 大陆直连实测 200（2026-10-21）。trae.ai / www.trae.ai / trae.cn / docs.trae.ai measured 200 direct from mainland (2026-10-21).
备案状态CAC filing: 备案未核验filing unknown：国内版所用模型的备案信息未核验到官方公示页（docs.trae.cn 为动态加载，本次未能直抓正文）；「未核验到」不等于「未备案」。no official filing disclosure page verified for the domestic edition's models (docs.trae.cn is dynamically loaded; body not fetched in the 2026-10-08 pass); "not verified" does not mean "not filed".

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: 官方文档（docs.trae.ai/ide/models）明示内置多模型并支持 API key 添加自定义模型；内置清单随版本变化（本次抓到的清单含 GPT 系/DeepSeek/Gemini 系等条目，未逐条清点为完整列表）。国际版（trae.ai）与国内版（trae.cn）模型面不同，以各自官方页为准。
The official docs (docs.trae.ai/ide/models) explicitly state built-in multi-model support, plus adding custom models via API key. The built-in list changes with versions (the list captured in the 2026-10-08 pass includes GPT-family / DeepSeek / Gemini-family entries; not enumerated as a complete list). The international edition (trae.ai) and domestic edition (trae.cn) differ in model surface — each official page is authoritative.
运行位置Where it runs: 官方口径三客户端：Desktop（独立应用，不依赖 TRAE IDE 运行，支持本地与云任务）、 Web（云端环境免安装）、Mobile（派发管理任务，设备离线自动切云执行）。 Desktop 给 Agent 窗口与 IDE 窗口双窗口；IDE 窗口内 IDE 模式与 SOLO 模式切换。
Official framing: three clients — Desktop (a standalone app that does not depend on TRAE IDE to run; supports local and cloud tasks), Web (a no-install cloud environment), Mobile (dispatch and manage tasks; auto-switches to cloud execution when the device is offline). Desktop gives dual windows: the Agent window and the IDE window; inside the IDE window you switch between IDE mode and SOLO mode.
本地文件Local files: 官方安全页明示：代码库文件始终留在本地设备（Privacy mode 语境）；沙箱执行把 agent 生成的命令放进受限环境（文件访问控制+高危命令拦截）。云任务线的数据边界未逐条核验。
The official security page states explicitly: codebase files always stay on the local device (in the Privacy mode context). Sandboxed execution puts agent-generated commands into a restricted environment (file-access control + high-risk command interception). Data boundaries of the cloud-task line: not verified one by one.
关机后的任务Tasks after shutdown: 后台/异步面是官方叙事的一部分：Desktop 支持长任务交云、Mobile 跨设备监控任务进度、离线自动切云执行——任务状态三端同步（官方文档口径）。
The background/async surface is part of the official narrative: Desktop supports handing long tasks to the cloud, Mobile monitors task progress across devices, offline auto-switches to cloud execution — task state syncs across the three clients (official docs framing).
工具与扩展Tools & extensions: agent 可搜代码库、多步计划、调工具；自定义 agent 可配 prompt/MCP 服务器/工具集（官方文档明示 MCP 配置面）； CUE 补全线（多行编辑/下一编辑预测）独立成页。
The agent can search the codebase, plan in multiple steps and call tools. Custom agents can configure prompt / MCP servers / tool sets (the official docs state the MCP configuration surface explicitly). The CUE completion line (multi-line edits / next-edit prediction) has its own page.
上下文与记忆Context & memory: 上下文可喂文件/文件夹/代码片段/终端输出/仓库/文档集/网页（官方文档列举）；跨会话记忆机制未核验到官方声明。
Context can be fed files / folders / code snippets / terminal output / repositories / document sets / web pages (enumerated in the official docs). Cross-session memory: not verified to an official statement.
权限与限制Permissions & limits: 账号体系（字节侧账号）；Privacy mode 开关（开启后官方声明不用你的聊天/代码/输出做分析、优化、训练）；沙箱执行开关。企业/团队管控面未逐条核验。
Account system (ByteDance-side accounts). Privacy mode toggle — when on, the official states your chats / code / outputs are not used for analysis, optimization or training. Sandboxed-execution toggle. Enterprise/team control surface: not verified one by one.
适合什么任务Best for: 适合：要桌面+Web+移动三端同一任务流、SOLO 模式整段委托、内置多模型免自己配 key 起步。不适合：要国内版国际版功能一致的预期（两条线）、要完全离线使用（云任务线是产品语义一部分）。
Good fit: desktop + Web + mobile sharing one task flow, SOLO mode for whole-segment delegation, built-in multi-model to start without configuring your own keys.

Not a fit: expecting the domestic and international editions to be feature-identical (they are two lines), wanting fully offline use (the cloud-task line is part of the product semantics).

头号误解Top misconception

国际版 trae.ai 与国内版 trae.cn 是两条线——模型清单与功能面别互相推断International trae.ai and domestic trae.cn are two lines — don't infer model lists or feature sets from one to the other
「代码不出本机」是 Privacy mode 语境的官方声明——云任务/Web/Mobile 线的数据边界要单独看（未逐条核验）'Code never leaves your machine' is the official claim in the Privacy-mode context — the data boundaries of cloud tasks / web / mobile lines need separate reading (not verified line by line)
SOLO 模式是「AI 主导全流程」（官方原文 AI takes the lead）——与 IDE 模式的控制权假设完全不同，按任务复杂度切SOLO mode means 'AI takes the lead' through the whole flow (official wording) — a completely different control assumption from IDE mode; switch by task complexity

未知项清单Known unknowns

国内版模型清单与备案公示位置（docs.trae.cn 动态加载未直抓）
云任务/Web 线的存储与处理边界（未逐条核验）
跨会话记忆机制（未核验到官方声明）
企业/团队管控面（未逐条核验）

The domestic edition's model list and where its filing is disclosed (docs.trae.cn dynamically loaded; not fetched directly)
Storage and processing boundaries of the cloud-task/Web lines (not verified one by one)
Cross-session memory mechanism (not verified to an official statement)
Enterprise/team control surface (not verified one by one)

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 2026-10-30，内容更新日content updated on 2026-10-21。

类型Type	名称Name	链接Link
official	TRAE 文档 · About TRAE（三客户端/双窗口/双模式/安全页）TRAE docs · About TRAE (three clients / dual windows / dual modes / security page)	https://docs.trae.ai/ide/what-is-trae
official	TRAE 文档 · Built-in models & custom modelsTRAE docs · Built-in models & custom models	https://docs.trae.ai/ide/models
changelog	TRAE 文档 · Changelog（changelog）TRAE docs · Changelog (changelog)	https://docs.trae.ai/ide/changelog
official	TRAE 国内版官网TRAE domestic edition official site	https://www.trae.cn

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（三条，都指向官方可读页没给的东西）：

国内版 trae.cn 的当前模型清单（docs.trae.cn 动态加载，本次未直抓到）
SOLO 模式高危命令在沙箱里的实际拦截行为
Mobile 离线自动切云的任务边界（哪些任务能切、上下文怎么带）

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

Suggested measurement (three, each pointing at something the official readable pages do not give):

The current model list of the domestic edition trae.cn (docs.trae.cn is dynamically loaded; not fetched directly the 2026-10-08 pass)
How SOLO-mode high-risk commands are actually intercepted inside the sandbox
Task boundaries of Mobile's offline auto-switch to cloud (which tasks can switch, how context carries over)