← Coding

DevinDevin

Cognition

闭源Closed source 云异步Cloud async 工单委派Issue delegation Slack集成Slack integration APIAPI partial lifecycle: active 核验Checked 2026-10-30

Devin 是「工单进、草稿 PR 出」的云异步软件工程师（Cognition）：会话在厂商云工作区跑，自带 shell/IDE/浏览器，Slack/Linear/Jira 直接派单，并行处理积压任务。

Devin is the 'ticket in, draft PR out' cloud async software engineer (Cognition): sessions run in the vendor's cloud workspace with its own shell/IDE/browser trio, dispatch work straight from Slack/Linear/Jira, and grind through backlogs in parallel.

适合与不适合When to use it, when not to

适合：批量委派工单、并行多任务、草稿 PR 再接管。不适合：要选模型、要全程本机执行、仓库里带密钥。

Fits: delegating backlog tickets in bulk, parallel multi-task work, @-dispatching from Slack, 'give me a draft PR then I take over'. Not a fit: picking your own models; everything running on your machine (Terminal only covers light tasks); repos carrying secrets.

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: 国外出品foreign-made　Cognition（美国）；与 Windsurf 同厂。Cognition (US); same vendor as Windsurf.
国内使用Direct access: 大陆可直连direct OK　devin.ai 与 docs.devin.ai 大陆直连实测 200（2026-10-21）。devin.ai and docs.devin.ai measured 200 direct from mainland (2026-10-21).
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（2026-10-21 检索）；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 (searched 2026-10-21); 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: 官方可读页（docs intro 与 release notes）未列「用户选模型」的面——模型与版本由厂商侧决定，这是厂商云形态的共性（本站判断，依据是「已查 intro 页未见 model 选择文档」，不是「官方声明不可选」）。
The official readable pages (docs intro and release notes) show no "user picks the model" surface — models and versions are decided on the vendor side. This is common to the vendor-cloud form (this site's judgement, based on "checked the intro page, no model-selection docs found" — not on "the official states selection is unavailable").
运行位置Where it runs: 会话跑在 Devin 的云工作区：网页端（app.devin.ai）内嵌三件套——Shell（看命令执行）、嵌入式 IDE（可实时接管改代码）、Browser（看它查文档/测应用）。另有 Devin for Terminal （cli.devin.ai 安装）在你本机跑，长任务用 /handoff 交给云端 Devin；还有 Devin API。
Sessions run in Devin's cloud workspace: the web app (app.devin.ai) embeds a three-piece set — Shell (watch commands execute), an embedded IDE (take over and edit code live), and a Browser (watch it read docs / test the app). There is also Devin for Terminal (installed from cli.devin.ai) running on your machine, with /handoff sending long tasks to the cloud Devin; plus the Devin API.
本地文件Local files: 云端会话不碰你本机文件——它在你授权的仓库上 clone、改、提 PR；仓库可见范围由集成授权决定。 Terminal 形态是例外：它直接在你本机环境跑，接管的是本地代码库。
Cloud sessions do not touch your local files — it clones, edits and opens PRs on repositories you authorize; repo visibility is set by integration grants. The Terminal form is the exception: it runs directly in your local environment, taking over the local codebase.
关机后的任务Tasks after shutdown: 主轴就是异步：官方按「自主 AI 软件工程师」定义它，适合并行处理多个任务——Linear/Jira 工单、整段新功能、bug 复现修复、迁移重构。官方经验法则原文：「if you can do it in three hours, Devin can most likely do it」（可验证性高的任务成功率高）。
The main axis is asynchrony: the official defines it as an "autonomous AI software engineer", suited to handling multiple tasks in parallel — Linear/Jira issues, whole new features, bug reproduction and fixes, migrations and refactors. The official rule of thumb (original wording): "if you can do it in three hours, Devin can most likely do it" (tasks with high verifiability have higher success rates).
工具与扩展Tools & extensions: 工作区自带 shell/浏览器/IDE 三类工具；与 Slack/Teams 线程直接 @Devin 派单，Linear/Jira 工单线集成。 MCP 支持状态未逐条核验（docs 有 integrations 分区，本次未逐页读）。
The workspace ships with three tool classes — shell, browser, IDE. Dispatch tasks by @Devin directly in Slack/Teams threads; Linear/Jira issue-line integrations exist. MCP support status: not verified one by one (docs have an integrations section; not read page by page the 2026-10-08 pass).
上下文与记忆Context & memory: 以会话为单位组织工作；知识沉淀/记忆面未逐条核验。官方建议把任务写清完成标准、拆成可验证步骤—— 上下文给多少由你在派单时决定。
Work is organized per session; the knowledge-accumulation/memory surface is not verified one by one. The official recommends writing tasks with clear completion criteria, broken into verifiable steps — how much context to give is decided when you dispatch.
权限与限制Permissions & limits: 账号+组织体系：企业管理员分配权限；仓库访问经集成授权（GitHub 等）划定范围；API 访问单独配。云端环境有联网能力（Browser 工具是产品语义的一部分），官方要求按对待公共计算面的标准看待它。
Account + organization system: enterprise admins assign permissions; repository access is scoped by integration grants (GitHub etc.); API access is configured separately. The cloud environment has internet access (the Browser tool is part of the product semantics) — the official asks you to treat it by the standard you would apply to a public compute surface.
适合什么任务Best for: 适合：把积压工单批量委派出去、并行跑多个任务、在 Slack 里随手 @ 派单、要「先给个草稿 PR 再接管」。不适合：要选模型、要全程本机执行（Terminal 只覆盖轻任务）、要把密钥放进被 clone 的仓库。
Good fit: delegating backlogged issues in bulk, running multiple tasks in parallel, @-dispatching casually in Slack, wanting "a draft PR first, then take over".

Not a fit: wanting model choice, wanting fully local execution end to end (Terminal only covers light tasks), keeping secrets inside the cloned repositories.

头号误解Top misconception

Devin 与 Windsurf 同属 Cognition，但是两条产品线——云异步工程师 vs IDE（Windsurf 档案另列），别混成一个Devin and Windsurf share vendor Cognition but are two product lines — cloud async engineer vs IDE (Windsurf profiled separately); don't merge them into one
「三小时法则」是官方经验法则不是保证——官方自己强调任务要「容易验证」（CI 能过/能自动部署验证）The 'three-hour rule' is the vendor's rule of thumb, not a guarantee — the official docs themselves stress tasks must be easy to verify (CI passes / auto-deploy checks)
Terminal 在本机跑、/handoff 把任务送去云端——同一工具里数据边界会切换，交接前先想清楚代码去哪Terminal runs on your machine, /handoff sends tasks to the cloud — the data boundary switches inside one tool; decide where your code goes before handing off

未知项清单Known unknowns

模型选择面是否存在（官方文档未见，未做实测）
MCP 支持状态（docs integrations 分区未逐页读）
知识/记忆沉淀机制细节（未逐条核验）
云端工作区的网络出口与持久化策略细节（未逐条核验）

Whether a model-choice surface exists (not found in official docs; not measured)
MCP support status (docs integrations section not read page by page)
Details of the knowledge/memory accumulation mechanism (not verified one by one)
Details of the cloud workspace's network egress and persistence policy (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	Devin Docs · Introducing Devin（产品定义/工作区三件套/接入方式）Devin Docs · Introducing Devin (product definition / workspace three-piece set / access paths)	https://docs.devin.ai/get-started/devin-intro
changelog	Devin Docs · Release Notes（changelog）Devin Docs · Release Notes (changelog)	https://docs.devin.ai/release-notes/overview
official	Cognition · Devin GA 博客Cognition · Devin GA blog	https://cognition.com/blog/devin-generally-available
official	Devin for Terminal · 安装入口Devin for Terminal · install entry	https://cli.devin.ai

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）：

模型面到底有没有选择——官方文档未见 model 选择页，实测登录后可确认
Slack 派单到 PR 开出的完整链路时长与失败重试行为
Terminal 与 /handoff 交接时，本地未提交改动怎么处理

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):

Whether a model-choice surface actually exists — no model-selection page found in the official docs; logging in and measuring can confirm
End-to-end latency and failure-retry behavior from Slack dispatch to PR opened
How uncommitted local changes are handled during Terminal ↔ /handoff transitions