接入 · 用 Connectors 让 Claude 读到你的数据Plug In · Letting Claude Read Your Data with Connectors.
你的上下文大多不在对话框里 —— 在 Gmail、Drive、Slack、Notion、Linear。Connectors（背后是 MCP 这个开放标准）让 Claude 直接读它们。这章讲哪些值得连、连上之后的边界，以及为什么「能连」不等于「该连」。Most of your context isn't in the chat box — it's in Gmail, Drive, Slack, Notion, Linear. Connectors (built on MCP, the open standard) let Claude read them directly. Which are worth connecting, the boundary once you do, and why 'can connect' isn't 'should connect.'
你的上下文大多不在对话框里。它在你的 Gmail、Drive、Slack、Notion、Linear 里 —— 你每次都得先去那边复制一段，再回到 Claude 粘上。Connectors 就是为了省掉这一趟。Most of your context isn't in the chat box. It's in your Gmail, Drive, Slack, Notion, Linear — and each time you go copy a chunk there and paste it back into Claude. Connectors exist to save that trip.
— I
Connectors 背后是 MCPConnectors Are MCP Underneath.
连接器让 Claude 直接读你别处的数据，不用你来回搬。Connectors let Claude read your data elsewhere, without you shuttling it back and forth.背后是 MCP，一个把 AI 接到「数据所在系统」的开放标准。1注 1Note 1Anthropic · Introducing the Model Context Protocol —— MCP 是连接 AI 与数据所在系统的开放标准，Connectors 即建立在它之上。Anthropic · Introducing the Model Context Protocol — MCP is the open standard for connecting AI to where data lives; Connectors are built on it.你不必懂协议 —— 在 claude.ai/connectors 浏览、连上一个应用，Claude 之后就能直接读它；3注 3Note 3Claude · Connectors 目录 —— 在 claude.ai/connectors 浏览可连接的应用，每个条目标明读写能力与可用范围。Claude · Connectors directory — browse connectable apps at claude.ai/connectors; each entry lists its read/write capabilities and availability.连接都在 Customize > Connectors 里管理。2注 2Note 2Claude 帮助中心 · 预置网页连接器（remote MCP）—— 网页连接器对所有用户开放；自定义连接器各档可用，Free 限 1 个；连接器可读可写（如建 Linear issue、发 Slack 消息），组织可限制写操作；在 Customize > Connectors 管理、断开或撤销权限。截至 2026-07。Claude Help Center · Pre-built web connectors (remote MCP) — web connectors are open to all users; custom connectors work on every plan, with Free limited to one; connectors read and write (e.g. open a Linear issue, send a Slack message), and organizations can restrict writes; managed, disconnected, or revoked under Customize > Connectors. As of 2026-07.Underneath is MCP, an open standard that wires AI to the systems where data lives.1注 1Note 1Anthropic · Introducing the Model Context Protocol —— MCP 是连接 AI 与数据所在系统的开放标准，Connectors 即建立在它之上。Anthropic · Introducing the Model Context Protocol — MCP is the open standard for connecting AI to where data lives; Connectors are built on it. You don't need the protocol — browse at claude.ai/connectors, connect an app, and Claude can read it directly afterward;3注 3Note 3Claude · Connectors 目录 —— 在 claude.ai/connectors 浏览可连接的应用，每个条目标明读写能力与可用范围。Claude · Connectors directory — browse connectable apps at claude.ai/connectors; each entry lists its read/write capabilities and availability. every connection is managed under Customize > Connectors.2注 2Note 2Claude 帮助中心 · 预置网页连接器（remote MCP）—— 网页连接器对所有用户开放；自定义连接器各档可用，Free 限 1 个；连接器可读可写（如建 Linear issue、发 Slack 消息），组织可限制写操作；在 Customize > Connectors 管理、断开或撤销权限。截至 2026-07。Claude Help Center · Pre-built web connectors (remote MCP) — web connectors are open to all users; custom connectors work on every plan, with Free limited to one; connectors read and write (e.g. open a Linear issue, send a Slack message), and organizations can restrict writes; managed, disconnected, or revoked under Customize > Connectors. As of 2026-07.顺手分清两类，免得混淆：你在 claude.ai 里连的是「网页连接器」（远程 MCP），跑在对方服务器上、所有客户端通用；另一类是「桌面扩展」（本地 MCP），装在 Claude 桌面端、跑在你这台机器上，能碰本地文件和工具，权限更深、也更该谨慎。这一章主要讲前者；后者用同一把尺子 ——「能只给读就别给写」—— 但风险更高。5注 5Note 5Claude 帮助中心 · 桌面扩展与网页连接器怎么选 —— 网页连接器是远程 MCP，跨客户端通用；桌面扩展是本地 MCP，装在 Claude 桌面端、跑在本机，能碰本地文件与工具。截至 2026-07。Claude Help Center · When to use desktop and web connectors — web connectors are remote MCP, shared across clients; desktop extensions are local MCP, installed in Claude Desktop and running on your machine, able to touch local files and tools. As of 2026-07.One pair worth keeping straight: what you connect inside claude.ai is a 'web connector' (remote MCP), running on the provider's servers and shared across all your clients; the other kind is a 'desktop extension' (local MCP), installed in Claude Desktop and running on your own machine, able to touch local files and tools — deeper access, more to be careful about. This chapter is mostly the former; the latter takes the same ruler — 'grant read-only when you can' — at higher stakes.5注 5Note 5Claude 帮助中心 · 桌面扩展与网页连接器怎么选 —— 网页连接器是远程 MCP，跨客户端通用；桌面扩展是本地 MCP，装在 Claude 桌面端、跑在本机，能碰本地文件与工具。截至 2026-07。Claude Help Center · When to use desktop and web connectors — web connectors are remote MCP, shared across clients; desktop extensions are local MCP, installed in Claude Desktop and running on your machine, able to touch local files and tools. As of 2026-07.连接器不只是只读。预置的网页连接器对所有用户开放；它既能读，也能写回 —— 让它在 Linear 建一个 issue、在 Slack 发条消息都行。组织可以限制写操作，比如只许读 Google Drive、不许它改文档。读还是读写，连之前就该看清。Connectors aren't read-only. The pre-built web connectors are open to all users; they read and also write back — have it open a Linear issue or send a Slack message. Organizations can restrict writes, say allowing it to read Google Drive but not edit documents. Read, or read-write — know which before you connect.连了能写的连接器，就别只让它「建议」。把动作说成命令，它才真去做 —— 这正是「说清」那套（第 2 章）：说该做什么，别问「能不能」。4注 4Note 4Claude API Docs · Prompting best practices —— 要它动手就把动作说成命令；别硬催工具（新版模型本就爱调工具，过度强调反而乱调）。截至 2026-07。Claude API Docs · Prompting best practices — phrase actions as commands to make it act; don't over-push tools (current models already lean toward tool calls, and over-emphasis misfires them). As of 2026-07.Once you've connected a connector that can write, don't let it merely 'suggest.' Phrase the action as a command and it actually acts — this is the 'say it clearly' craft (Chapter 2): state what to do, don't ask 'could you.'4注 4Note 4Claude API Docs · Prompting best practices —— 要它动手就把动作说成命令；别硬催工具（新版模型本就爱调工具，过度强调反而乱调）。截至 2026-07。Claude API Docs · Prompting best practices — phrase actions as commands to make it act; don't over-push tools (current models already lean toward tool calls, and over-emphasis misfires them). As of 2026-07.
只是建议Just a suggestion
你能不能在 Linear 里给这个 bug 建个 issue？Could you maybe open a Linear issue for this bug?
直接动手Just do it
在 Linear 的 Web 项目里建一个 issue：标题用这条报错，正文贴复现步骤，优先级 High。Open an issue in the Linear Web project: title it with this error, paste the repro steps in the body, priority High.
提示词PromptGitHub 连接器 · 问代码库GitHub connector · ask a repo
连的是 <owner/repo> 这个仓库。
我要弄清：<一个具体问题，比如「登录超时是哪个改动引入的」>。
先读相关文件和最近的提交，没读过的别下结论。
给我：<涉及的文件 + 关键行 + 一句话原因>，附上你看的 commit 或文件链接。
只读：别建 issue、别改任何东西，除非我明说。The repo is <owner/repo>.
I need to pin down: <a concrete question, e.g. which change introduced the login timeout>.
Read the relevant files and recent commits first; don't conclude about code you haven't opened.
Give me: <the files + the key lines + a one-line cause>, with links to the commits or files you read.
Read-only: don't open an issue or change anything unless I say so.
连的是 acme/web 这个仓库。
我要弄清：结账页偶发的 500 是哪个改动引入的，范围最近两周。
先读 src/checkout 下相关文件和最近的提交，没读过的别下结论。
给我：涉及的文件 + 关键行 + 一句话原因，附上你看的 commit 链接。
只读：别建 issue、别改任何东西，除非我明说。The repo is acme/web.
I need to pin down: which change introduced the intermittent 500 on the checkout page, over the last two weeks.
Read the relevant files under src/checkout and the recent commits first; don't conclude about code you haven't opened.
Give me: the files + the key lines + a one-line cause, with links to the commits you read.
Read-only: don't open an issue or change anything unless I say so.
提示词PromptGoogle Workspace · 跨邮件与云盘问Google Workspace · ask across mail & drive
在我的 Gmail 和 Drive 里找：<一个跨来源的问题，比如「某客户这季度答应过什么」>。
范围：限定<时间段 / 发件人 / 文件夹>，别翻无关的。
给我：<一句话结论 + 出处（邮件主题或文件名）>，每条能点回原件。
要写邮件就只起草、存草稿，别发；发不发我自己来。In my Gmail and Drive, find: <a cross-source question, e.g. what a client promised this quarter>.
Scope: limit to <a date range / sender / folder>; don't trawl the unrelated.
Give me: <a one-line answer + the source (email subject or file name)>, each clickable back to the original.
If an email is needed, only draft it and save the draft — don't send; sending is mine.
在我的 Gmail 和 Drive 里找：Acme 这个客户在 Q2 的邮件和会议纪要里答应过哪些交付。
范围：限定 4–6 月、与 @acme.com 往来、以及 Drive 里「Acme」文件夹，别翻无关的。
给我：一句话结论 + 出处（邮件主题或文件名），每条能点回原件。
顺手把跟进邮件起草好、存成草稿，别发；发不发我自己来。In my Gmail and Drive, find: what deliverables the client Acme promised across Q2 emails and meeting notes.
Scope: limit to April–June, correspondence with @acme.com, and the 'Acme' folder in Drive; don't trawl the unrelated.
Give me: a one-line answer + the source (email subject or file name), each clickable back to the original.
Draft the follow-up email and save it as a draft — don't send; sending is mine.
— II
能连不等于该连Can Connect Isn't Should Connect.
连接器的清单很长，值得你连的没几个。The connector list is long; the ones worth connecting are few.每连一个数据源，就多一份权限、多一处它能看到的东西。2注 2Note 2Claude 帮助中心 · 预置网页连接器（remote MCP）—— 网页连接器对所有用户开放；自定义连接器各档可用，Free 限 1 个；连接器可读可写（如建 Linear issue、发 Slack 消息），组织可限制写操作；在 Customize > Connectors 管理、断开或撤销权限。截至 2026-07。Claude Help Center · Pre-built web connectors (remote MCP) — web connectors are open to all users; custom connectors work on every plan, with Free limited to one; connectors read and write (e.g. open a Linear issue, send a Slack message), and organizations can restrict writes; managed, disconnected, or revoked under Customize > Connectors. As of 2026-07.自定义连接器各档都能用，但 Free 只能连 1 个 —— 这反而是个好默认：先连你每天都翻、且最信任的那一两个（多半是 Gmail / Google Drive，或 Slack、Linear 这类你每天开着的），别一上来把目录连满。Every data source you connect adds a permission and one more place it can see.2注 2Note 2Claude 帮助中心 · 预置网页连接器（remote MCP）—— 网页连接器对所有用户开放；自定义连接器各档可用，Free 限 1 个；连接器可读可写（如建 Linear issue、发 Slack 消息），组织可限制写操作；在 Customize > Connectors 管理、断开或撤销权限。截至 2026-07。Claude Help Center · Pre-built web connectors (remote MCP) — web connectors are open to all users; custom connectors work on every plan, with Free limited to one; connectors read and write (e.g. open a Linear issue, send a Slack message), and organizations can restrict writes; managed, disconnected, or revoked under Customize > Connectors. As of 2026-07. Custom connectors work on every plan, but Free allows just one — which is a good default: connect the one or two you check daily and trust most (usually Gmail / Google Drive, or the Slack and Linear you keep open all day), don't fill the directory on day one.连之前先想清楚它能看到什么、能改什么；连之后，权限不是一锤子买卖 —— 在 Customize > Connectors 里随时能看每个连接器的权限、改设置、或直接断开。能连进来的多，该连的少，撤起来要方便。Before connecting, be clear what it can see and change; after, the permission isn't a one-time deal — under Customize > Connectors you can review each connector's access, change settings, or disconnect at any time. Plenty can connect, few should, and revoking should stay easy.同一份资料，进 Claude 有好几条路，按「多久要一次、要不要回写」选：The same material can reach Claude several ways; choose by 'how often you need it' and 'whether it writes back':
你的资料Your material
走这条Route
为什么Why
这次用一下的几个文件A few files for this one time
直接上传到对话Upload to the chat
一次性，用完就算，不必留存One-off; done when done, nothing to keep
一件长期事反复引用的固定材料Fixed material a long-term thing cites again and again
Project 知识库（第 4 章）Project knowledge base (Ch. 4)
钉在工作区、会被检索，不必每次重传Pinned, retrieved on demand, no re-upload
活的、常变的数据（邮件、任务、文档）Live, changing data (mail, tasks, docs)
连接器（远程 MCP）Connector (remote MCP)
Claude 直接读源头，还能回写Claude reads the source directly, and can write back
本机的文件与工具Local files and tools
桌面扩展（本地 MCP）Desktop extension (local MCP)
跑在你机器上，权限最深、最该谨慎Runs on your machine — deepest access, most caution
— III
互动式连接器Interactive Connectors.
有些连接器不只是读数据，还能在 Claude 里开出一个能点的界面。Some connectors don't just read data — they open a clickable interface inside Claude.有些连接器是互动式的：它能在对话里直接渲染出活的界面 —— 仪表盘、任务板、设计工具。不只是把数据读进来，还能让你在对话里挑选、确认、提交。Some connectors are interactive: they render a live interface right in the conversation — dashboards, task boards, design tools. Not just reading data in, but letting you select, confirm, and submit inside the chat.哪类活适合互动式：要在外部系统里挑选、确认、提交的操作；哪类纯读取就够：你只想把那边的内容拿进来参考。再往深处，自定义连接器要接你公司内部的 API —— 那已经越过普通用户、进了开发者地界，这本书不展开。Which work suits interactive: actions where you select, confirm, and submit in an external system; which is fine read-only: you just want that content pulled in for reference. Deeper still, a custom connector wiring your company's internal API crosses out of ordinary-user territory into developer land — beyond this book.把这章用到你工作里最常翻的数据源上：Apply this to the data sources you check most:
列出你工作里最常翻的 3 个数据源（邮箱 / 云盘 / Slack / 任务系统）。List the 3 data sources you check most at work (email / drive / Slack / task tracker).
只连其中你最信任的 1 个，别贪多。Connect only the one you trust most — don't overreach.
问一个必须跨越「对话 + 那个源」才答得了的真实问题。Ask a real question that can only be answered by spanning the chat plus that source.
记录它替你省了几次切窗口、复制粘贴；再回 Customize > Connectors 看一眼它现在能看到你哪些东西。Note how many window-switches and copy-pastes it saved — then check under Customize > Connectors what it can now see of yours.
能连进来的数据很多，
该连的没几个.
Plenty of data can connect,
few should.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return