What Are Claude Connectors? What Claude Can See

What are Claude connectors? They're authorized links between Claude and the individual apps and services you already use — your Google Drive, Linear, Slack or database — that let Claude retrieve your data and take actions inside those services on your behalf. Each connector covers one service, you add only the ones you need, and Claude can never reach more than your own account already can. This article explains how they work and, more importantly, exactly what Claude can and can't see once you connect one.

• A connector is one authorized link between Claude and one external service — one connector, one service.
• Every connector runs on an MCP server (Model Context Protocol, an open standard for exposing tools and data to AI clients), but "connector" and "MCP server" describe different layers of the same thing.
• Claude inherits your permissions. If you can't open a file, channel or record in the source app, the connector can't reach it either.
• Access is capped three ways: your permissions in the service, the tools the developer built, and the scopes you approve at sign-in.
• Read and write are separate, and write actions are typically gated behind a per-action confirmation that's on by default.
• ChatGPT and Cursor offer the same capability under different names and setups, because all three converged on MCP.

What Are Claude Connectors?

Think of Claude the way you'd think of a phone fresh out of the box. It does a lot on its own — it can reason, write, summarize and code — but it can't see the things that live in your other accounts. It doesn't know what's on your project board, what's in your files, or what last quarter's numbers were. To reach any of that, you install an app. On Claude, that app is called a connector.

So what are Claude connectors, exactly? A connector is an authorized link between Claude and one specific external service — your Google Drive, your Linear workspace, your Slack, your database — that lets Claude retrieve data from that service and take actions inside it. Anthropic's own description is direct: connectors "let Claude access your apps and services, retrieve your data, and take actions within connected services." One connector, one service. You add the ones you need and leave the rest alone.

Under the hood, a Claude connector is backed by an MCP server. MCP — the Model Context Protocol — is an open standard for exposing external tools and data to AI clients in a structured way, and a connector is Anthropic's product name for an MCP-backed integration you've authorized. In practice you can treat the two words as two layers of the same thing: connector is what you click on in Claude's interface; MCP server is the software on the other end that actually talks to the service. (Section 6 goes deeper on whether they're "the same thing.")

What changes once you connect one

Here's the shift in a single sentence: once a connector is authorized, Claude can call that service's tools directly inside a conversation, instead of you copying and pasting data in by hand.

That difference is easy to underrate until you feel it. Without a connector, working with your own data means switching to Google Drive, finding the document, copying the relevant chunk, pasting it into the chat, and repeating that for every file. With a Google Drive connector, you can ask Claude to search a folder, pull the right documents and summarize them — it fetches what it needs and works with it in place. The same pattern applies elsewhere: connect Linear and Claude can create issues; connect Slack and it can send a message; connect a CRM and it can look up a record.

The vocabulary, kept straight

• Connector — Claude's name for an authorized integration with one external service.
• MCP server — the standard-based software that a connector talks to; it defines the tools and data Claude can reach.
• Tool — a single action a connector exposes, such as "search files" or "create issue." Claude decides when to call one based on your request.

One more thing worth flagging early, because it's the whole point of this article: a connector doesn't hand Claude the keys to everything in the connected service. Claude inherits your permissions. If you can't open a particular file, channel or record in the source system, the connector can't reach it through Claude either. The rest of this piece walks through exactly what that means for the data Claude can — and can't — see.

What Data Can Claude Access Through Connectors?

Claude can only reach what three things allow: your own permissions in the connected service, the specific tools the connector's developer built, and the scopes you granted when you authenticated. In practice this means connecting a service never hands Claude a complete copy of everything in that app — it gets scoped, on-demand access to the same things you can already reach yourself.

The clearest way to understand claude connector data access is Anthropic's own rule: Claude inherits each person's permissions from the connected service. If you can't open a particular file, channel, or record in the source system, the connector can't reach it from Claude either. Connecting Google Drive doesn't let Claude read documents you were never shared on; connecting Slack doesn't expose private channels you aren't a member of.

The three limits, in order

• Your permissions in the service. A connector operates as you. Access controls in the underlying app — file sharing, channel membership, CRM record ownership — carry straight through.
• The tools the developer built. A connector exposes a fixed set of actions (a "tool" is a single named operation the connector offers, like "search files" or "create issue"). If the developer only built read operations, Claude can't write, no matter how you ask.
• The scopes you granted at sign-in. During authentication you approve a set of permissions (scopes). Claude's reach is capped by what you approved — a read-only scope stays read-only.

So when people ask what data can Claude access through connectors, the honest answer is: a narrow, permissioned slice, requested tool-by-tool as the conversation needs it — not a bulk export.

What about Team and Enterprise plans?

The same model holds on Team and Enterprise plans. Each person still connects and authenticates individually, so one colleague's connection to a shared service reflects that colleague's own permissions — it doesn't widen access for everyone. The exception is Anthropic's separate enterprise-managed authentication, currently in beta, where an administrator authorizes access centrally rather than each person authenticating on their own. Even there, the underlying service's own access controls still apply.

One practical note worth checking before you connect: each connector's page in the directory documents its read/write capabilities. Reviewing that — and the scopes shown on the authentication screen — tells you exactly what a connector can and can't touch before you approve it.

How Do Claude Connectors Work?

Under the hood, a connector is an MCP server — a small service that speaks the Model Context Protocol, an open standard for how AI assistants talk to external tools and data. Claude is the client: the program that connects to that server, reads what it offers, and decides when to use it. So when people ask how do Claude connectors work, the short answer is that Claude and the server have a structured conversation in the background while you have a normal one in the chat window.

An MCP server exposes its capabilities as a few distinct primitives:

• MCP tools — actions Claude can call, like "search files," "create issue," or "send message." Each tool has a name, a description, and a defined set of inputs. Claude picks a tool and fills in the arguments; the server runs it and returns structured data.
• MCP resources — data Claude can read, such as a document, a record, or a file's contents. Resources are pulled in as context rather than triggering an action.
• Prompts — a third primitive: pre-written templates the server can offer. Most connectors don't surface these directly in Claude today, so you'll mostly deal with tools and resources.

What happens when you ask a question

When your question needs live data, the flow looks like this:

1. Claude reads the list of tools and resources the connector advertises.
2. It decides one of them is relevant — for example, calling a "search documents" tool with the terms from your question.
3. The server runs the call against your account and returns structured data (not a screenshot or a guess — actual fields and values).
4. Claude writes its reply from that returned data.

This is why a connector's answers reflect your real, current information rather than the model's training data. It's also why, as one commentator put it, connectors "burn tokens before you type" — the tool descriptions and returned data both count as context the model has to read.

The consent screen is your moment of control

Connecting a service for the first time works much like installing a phone app. Before anything is linked, you're sent to the service's own sign-in page and shown an OAuth consent screen — OAuth 2.1 is the standard that lets you grant an app limited access without handing over your password. The screen lists exactly what Claude is asking to do: read your files, create issues, send messages. It's the same prompt as an app asking "Allow access to your photos?" You review the requested permissions, then approve or decline.

That approval is the whole security boundary in one click. Nothing connects until you say yes, and — per Anthropic's help center — Claude inherits your permissions from the connected service. If you can't open a specific file, channel, or record in the source app, the connector can't reach it from Claude either. Approving a connector does not grant Claude more access than you already have.

Where connectors run, and on which plans

According to Anthropic's help center, connectors work across Claude, Claude Desktop, Claude Code, and the API (via the MCP Connector), and web connectors are available across Claude, Cowork, Claude Desktop, and Claude Mobile on iOS and Android. Plan availability and any per-plan limits — including how many connectors free-plan users can add — change over time, so check the current details on Anthropic's help center before you rely on them (checked 2026-08-16).

The Permission Model in Plain English

There is no single rule that says "Claude can read this but not that." What Claude can do through a connector — a one-click bridge between Claude and an app you already use — depends on how that specific connector was built and which tools it exposes. A tool, in Model Context Protocol (MCP) terms, is a single named action the connector offers to Claude, like "search files" or "create issue." Two connectors for two different services can behave very differently, so it pays to read each connector's own page before you trust it.

Access follows your own account boundary

For connectors where you authenticate as yourself — you log in to the service and grant access — Claude generally can't see past your own account boundary in that service. Anthropic states it plainly: Claude inherits each person's permissions from the connected service, so if you can't open a specific file, channel, or record in the source system, the connector can't reach it from Claude either. Connector scopes (the specific permissions you agree to during sign-in) define the outer edge of what's possible; your account's own access defines what's actually reachable inside that edge.

Read and write are separate — and both depend on the connector

It helps to treat read and write as two independent questions:

• Read means Claude can retrieve and look at your data — search documents, pull records, summarize a folder.
• Write means Claude can change something in the service — create an issue, send a message, update a record.

A connector might expose only read tools, in which case Claude can look but never change anything. Another might expose write tools as well. Each connector's directory page documents its read/write capabilities, so you can check before connecting rather than guessing.

Per-action confirmation on write

When a connector can take actions, write steps are typically gated behind a confirmation prompt that is on by default: Claude tells you what it's about to do, and the action runs only after you approve it. This is a core piece of the security and privacy model — it keeps a write from happening silently in the middle of a longer task.

On Team and Enterprise plans, a workspace owner can choose to let members skip that per-action confirmation for a given connector. That trades a bit of friction for speed, and it's a deliberate decision: once confirmation is off, an approved write tool can run without a per-action check. Turn it off only for connectors and actions you already trust.

An allowed write still needs your underlying permission

One more limit is easy to miss: even when a write tool is exposed and confirmation is satisfied, the action still has to clear your own permissions in the source system. If your account can't post to a particular Slack channel or edit a specific record, Claude can't do it on your behalf either — the connector can only ever act within the rights you already hold. The connector doesn't grant new powers; it lets Claude use the ones you have.

A Data-Access Example: Four B77 Apps Side by Side

Abstract permission talk gets clearer with real examples. The table below shows four connectors available on B77 and exactly what each one lets Claude see and do. The pattern to notice: two of the four are read-only by design — Claude can look but never change anything — while the other two can write, but only after asking you to approve each action.

These are useful claude connector examples for thinking about the difference between a read only vs read write connector before you connect anything.

Reading the table

Search Console and Web Analytics are read-only. That's a deliberate design choice: their value is in answering questions ("which queries lost impressions last week?", "where do people drop out of the signup funnel?"), and answering questions never requires the ability to change your setup. Even if a prompt asked Claude to modify tracking or verification, the connector simply has no tool for it.

Google Ads and Knowledge Base can write, but neither writes silently. When Claude proposes creating a campaign, adjusting a budget or updating a document, the action is surfaced to you for approval before it runs. You stay the one who clicks "go."

Two limits carry across every write-capable connector here. First, the connector inherits your permissions — it acts as you, so it can never do something your own account role doesn't allow. Second, it can't reach data outside what's shared with you, so a Knowledge Base connector won't surface a project you were never given access to.

Are Claude Connectors the Same as MCP?

Short answer: every Claude connector runs on MCP, but the two words are not interchangeable. One is a protocol; the other is a product feature that uses it.

MCP — the Model Context Protocol — is the open standard Anthropic published in November 2024 for connecting AI clients (like Claude or a coding tool) to external tools and data sources. It defines how a client asks a server what it can do, calls a "tool" (an action the server exposes), and reads a "resource" (data the server makes available). MCP is a spec, not a product: anyone can implement it, and servers built to it can talk to any client that speaks the same protocol.

A connector is Claude's product-level name for an authorized, wired-up MCP-backed integration living inside your account or organization. When you connect Claude to Google Drive or Linear and grant it access, you've created a connector. Under the hood it's an MCP server that Claude has been authorized to reach; "connector" is just the label Claude puts on that finished, ready-to-use link.

So when people ask about claude connectors vs mcp, the cleanest way to hold it in your head is:

• MCP — the open protocol underneath (the "how").
• MCP server — a running service that implements the protocol and exposes tools and resources.
• Connector — Claude's word for an MCP server you've connected and authorized in your account.

The reason it matters to keep these separate: "MCP server" is the broader term. The same server that shows up as a connector in Claude can be wired into other assistants and coding tools that support the protocol — the server doesn't know or care that Claude calls it a "connector." That's the whole point of an open standard: build once, connect to many clients.

So the answer to are Claude connectors MCP is yes, in the sense that every connector is backed by an MCP server — but not every MCP server is a Claude connector, and MCP itself is a much bigger thing than any single client's feature.

If you want the protocol explained from the ground up — tools, resources, transports, and how a client and server actually exchange messages — see our companion explainer on MCP. This article stays focused on connectors as you experience them inside Claude.

Does ChatGPT Have Connectors Like Claude? (And Cursor?)

Short answer: yes, the same underlying plumbing exists, but each product names and packages it differently. If you're asking does ChatGPT have connectors like Claude, the useful thing to understand is that all three—Claude, ChatGPT and Cursor—now speak the same protocol underneath: MCP (Model Context Protocol), the open standard that lets an AI assistant call external tools and read external data through a defined interface. Once you grasp that, the differences are mostly about the user interface and the plans each vendor offers.

ChatGPT: "apps" as the umbrella term

OpenAI groups this capability under a single category it calls apps, folding what you might think of as data connectors and interactive, in-chat apps into one directory rather than keeping them as separate concepts. In practice an "app" in ChatGPT can both retrieve your data from a service and present interactive elements inside the conversation. Underneath, these connect through remote MCP servers—the same server type you would point Claude at.

Two things are worth checking against OpenAI's own documentation before you rely on them, because they move and because plan availability changes:

• Remote MCP support. Confirm the current state of remote MCP server support in the developer tooling and in the ChatGPT app directory directly on OpenAI's docs, since the rollout has shifted over time.
• Read vs. write by plan. Whether an app can only read your data or can also take actions (write), and which subscription tiers get which apps, is set per plan. Read the plan comparison on OpenAI's site rather than trusting a third-party summary.

The practical takeaway: the vocabulary differs—Claude says "connectors," ChatGPT says "apps"—but the standard converged, so a well-built MCP server can generally serve both.

Cursor: a narrower, developer-first setup

Cursor, the AI code editor, exposes MCP through configuration rather than a consumer-style directory. You add servers either in a project-level config file or in Cursor's global settings, and it supports both local stdio servers (a process running on your own machine that Cursor talks to over standard input/output) and remote servers over SSE (server-sent events, a one-way streaming transport) or streamable HTTP (the newer HTTP-based transport in the MCP spec). Cursor also caps the total number of active tools it will load at once, so if you connect several servers you may need to disable some to stay under the limit. This is a developer-oriented workflow: more manual, more control, less hand-holding than a click-to-connect directory. Check Cursor's docs for the current config format and the exact tool cap, as both have changed between releases.

MCP clients comparison at a glance

One naming aside, since people search for it: Claude connectors vs plugins. "Plugin" is an older, vendor-specific term from an earlier era of AI extensions. Connectors and apps are the current, MCP-based way in; if a source is still talking about "plugins," treat it as potentially out of date and confirm against the vendor's live documentation.

When Not to Connect Something

Connecting a service to Claude is convenient, but every connection is a standing grant of access — an OAuth 2.1 authorization (the industry-standard way an app gets permission to act on your behalf without seeing your password) that stays live until you revoke it. Knowing when not to connect is as useful as knowing how. Here are the cases where you should pause, and the safer alternative for each.

• The app holds sensitive data Claude doesn't need to act on. Health records, financial accounts, and unencrypted personal data (PII) are poor candidates for a live connection. If you only want Claude to reason over the contents once, export the relevant records to a file and paste or upload that instead. A read-once export cannot be queried again later; a live connector can.
• The task is a one-off. If you need something done a single time, a persistent OAuth grant sitting around unused is one of the quieter Claude connector risks — it's access you're no longer watching. For one-shot work, prefer a manual export, or connect, finish the task, and revoke the grant immediately afterward.
• The requested scopes outgrow the task. Scopes define what a connector can do — read, write, delete, and which parts of your account. If a server asks for write or admin access to summarize a few documents, that's a mismatch. Look for a narrower or read-only option before you approve, and decline if the server won't grant less.
• The operator of the server is unclear. An MCP server runs code that receives whatever you connect to it. Review an unfamiliar server the way you'd review any new third-party app: who operates it, what it stores, where it's hosted, and how to revoke access. If you can't answer those questions, don't hand it real data — test with a throwaway or non-sensitive account first.

The common thread: match the level of access to what the task genuinely requires, and treat a live connection as an ongoing relationship rather than a one-time click. When in doubt, read from an export, choose the read-only option, or revoke the grant once you're done.

Frequently Asked Questions

What data can Claude access through connectors?

A connector lets Claude read your data and take actions inside a connected app or service — for example searching files in Google Drive, creating issues in Linear, or sending messages in Slack. Crucially, Claude inherits your permissions from the source system: if you can't open a specific file, channel, or record there, the connector can't reach it either. So "what are Claude connectors used for" comes down to bringing your existing tools into a chat, within the same access limits you already have.

Are Claude connectors the same as MCP?

They're closely related but not identical. MCP (Model Context Protocol) is the open standard that defines how an AI assistant talks to an external tool; a Claude connector is Anthropic's user-facing wrapper for connecting to a service — often, though not always, one that speaks MCP. In short, most connectors are MCP under the hood, but "connector" describes the product experience while MCP describes the protocol.

Are Claude connectors the same as Claude integrations?

In everyday use the terms overlap, and Anthropic's documentation treats connectors as the way you integrate Claude with your apps and services. If you see "Claude integrations" in older articles or menus, it's referring to the same idea of linking Claude to an external tool. The current, documented name in the Claude interface is connectors, and they work across Claude, Claude Desktop, Claude Code, and the API via the MCP Connector.

Does ChatGPT have connectors like Claude?

Yes — ChatGPT offers its own connectors to bring in data from external apps, and both assistants increasingly build on MCP as the shared standard. The setup steps, available services, and naming differ between the two products, so a connector you enable in one won't automatically appear in the other. If you're comparing them, check each vendor's current documentation, since support changes quickly (this answer reflects the state as of 2026).

Is there a full list of Claude connectors?

Anthropic maintains a built-in Connectors Directory inside Claude, Claude Desktop, and the mobile apps, where each entry lists its use cases and read/write capabilities. For a curated view of ready-made, hosted options you can also browse B77's directory of Claude connectors. Note that a connector doesn't necessarily add cost on Claude's side, but it will consume tokens whenever Claude uses it, so "are Claude connectors free" depends on both your Claude plan and any subscription for the underlying app.

Once a couple of connectors are in place, they compound: see how a product manager turns raw feedback, research notes and tasks into AI-assisted roadmap building in our workflow guide.

The bottom line

The reassuring thing about connectors is that they don't expand what Claude can see — they just save you the copy-and-paste. Access is capped three ways every time: by your own permissions in the service, by the tools the connector's developer built, and by the scopes you approve at sign-in. Read that authentication screen, prefer read-only where it's enough, keep write confirmation on for anything you don't fully trust, and revoke grants you're no longer using. When you're ready to add one, you can browse hosted, ready-to-connect MCP servers — with their read/write capabilities documented up front — in the B77 directory and connect them to Claude in a couple of clicks, with auth handled for you.Browse hosted MCP servers and connect one to your assistant in a couple of clicks: the B77 marketplace.