Claude connectors: how they work and how to add a custom one

A Claude connector lets Claude read from and act in another app. Connect Linear and Claude can open an issue. Connect Google Drive and it can search your files. Every connector is an MCP server underneath, which is why the same connector works in Claude on the web, Claude Desktop, the mobile apps, Claude Code and the API.

This guide covers what connectors can reach, which plans get them, how to add one from the directory or by URL, and how to limit what each one may do. We checked every step against Claude's settings on October 28, 2026, while adding Corsair's own MCP server as a custom connector, so the screens described here are the current ones.

What a Claude connector is

A connector is an MCP server that Claude talks to on your behalf. The server offers tools, such as "search files" or "create issue", and Claude decides when to call them based on what you ask. You usually sign in to the app once, during setup, and Claude uses that sign-in from then on.

There are three kinds:

• Directory connectors. Listed in Claude's Connectors Directory and built by the app maker or by Anthropic. Google Drive, Gmail, Slack, Notion and Figma are all here. They run on the provider's servers.
• Custom connectors. Any remote MCP server you add by its URL. Use them for apps without a directory listing, internal tools, or a server that covers many apps at once. Claude reaches them from Anthropic's cloud.
• Desktop extensions. A .mcpb file you install in Claude Desktop under Settings > Extensions. It runs on your own computer, which suits local files, local tools and offline work.

Once a connector is added, Claude can bring it into a conversation on its own when it fits the request. You can also switch individual connectors on or off for each chat from the + menu, under Connectors.

Which plans get connectors

Every plan does, including Free. The differences are in custom connectors and in who may add them.

• Free. Directory connectors, plus one custom connector.
• Pro and Max. Directory connectors and custom connectors, added to your own account.
• Team and Enterprise. An Owner or Primary Owner turns a connector on for the organization, then each member signs in with their own account. On Team, only Owners add custom connectors. On Enterprise, people with a custom role that can manage the organization's libraries can add them too.

Members who can't add a connector can ask for one with the Request button on its directory page. Requests are on by default for Team and off for Enterprise. Enterprise also has Enterprise-managed auth, in beta, where an admin authorizes a connector once and everyone inherits access on first login.

The Free limit matters more than it looks. One custom connector means one MCP server. If you need several apps that aren't in the directory, a server that covers many apps behind one URL is the only way to get all of them on Free. That's the gap Corsair fills, which we come back to below.

What data Claude can access through a connector

Settle this before you connect anything with write access. The answer is narrower than many people expect.

• Only what your account can see. Claude inherits the permissions of the account you connect. If you can't open a file, channel or record in the app, Claude can't reach it through the connector either.
• Shared credentials are the exception. A custom connector set up with one shared credential, such as an API key sent as a request header, reaches whatever that credential can reach, for everyone who uses it.
• Only when the connector is on. Claude uses a connector in a chat only if it's switched on for that conversation.
• The app processes the data. Connected services handle data on their own servers, under their own terms. Claude's own data settings, such as US-only inference on Enterprise, don't change where a third-party service runs.
• Team and Enterprise limits. Connectors work only in private projects, and chats that contain synced content can't be shared.

Data is encrypted in transit. The bigger practical risk is prompt injection, where a document or email contains instructions aimed at Claude. Claude has protections against it, but a connector with write access is where a successful injection would do damage. That's why the next section matters.

How to control what a connector can do

Claude groups each connector's tools by type, such as read-only tools and write/delete tools. You set each group, or each tool, to one of three states:

• Always allow. Claude calls the tool without asking.
• Needs approval. Claude asks you before every call.
• Blocked. Claude can't use the tool.

Open Customize > Connectors and select the connector to see its Tool permissions. On Team and Enterprise, Owners can set these for the whole organization, for example letting Claude read and summarize email but never send it. Members can't override an organization-wide setting.

Once you have many connectors, check one more setting. Under the + menu, Connectors > Tool access, the default is Auto, which works for most people. Anthropic suggests On demand once 10 or more connectors are active, to leave more room in the conversation.

How to add a connector from the directory

1. Open Customize > Connectors in Claude, or choose Connectors > Manage connectors from the + menu in a chat.
2. Browse by category or search, then open the connector's page. It lists what the connector can read and write.
3. Select Connect and sign in to the app when it asks.
4. In a chat, check that the connector is switched on under the + menu.

On Team and Enterprise, an Owner adds it first from Organization settings > Connectors > Browse connectors. Members then connect with their own accounts.

How to add a custom connector by URL

You need the URL of a remote MCP server, which often ends in /mcp. The server has to be reachable from the public internet, because Claude connects to it from Anthropic's cloud, not from your computer. That holds in Claude Desktop too.

1. Open Customize > Connectors, select Add, then Add custom connector.
2. Enter a name, which is what you'll see in your connectors list, and the MCP server URL. Select Continue.
3. Claude checks the server. It connects, looks up the server's sign-in settings and checks the sign-in provider. A wrong URL fails at the first check with "Not found 404".
4. The second step shows what Claude found and marks the options it detected. In most cases, keep them.
5. Select Add, then Connect, and sign in.

The options on the second step are worth understanding, because you can't change them later. To switch, you remove the connector and add it again.

Authentication decides when people sign in:

• Sign in now. Each user signs in through the server's OAuth flow before any tool works.
• Sign in when needed. Claude connects without credentials and asks you to sign in when the server requires it. This suits servers that offer some tools without an account.
• No sign-in. For open servers, or servers that take an API key, which you add under Request headers.

OAuth client decides how Claude identifies itself to the server:

• Use Claude's published identity. Anthropic's recommended option. The server reads Claude's details from a Client ID Metadata Document, or CIMD, that Anthropic hosts, so there's nothing to set up. The server has to support it.
• Register automatically. Claude registers itself with the server as people connect, using Dynamic Client Registration, or DCR. Most MCP servers support this.
• Use your own OAuth client. For servers where you've already registered a client ID.

When we added Corsair's server, Claude detected Sign in now and Register automatically. We kept both, and it connected on the first try. That's the safe default for any server. Claude reads the server's sign-in settings before you choose, so what it marks as detected is what the server supports.

Request headers let you send up to four fixed headers, such as an API key, with every request. Anthropic says this is in beta for a limited set of organizations, so you may not see it.

If Claude says "A connector with this URL already exists in your organization", someone has already added that server. Use the existing connector instead of adding a second one.

Interactive connectors

Some connectors show a live interface in the chat instead of text. That can be an inline card for a quick action, or a fullscreen view for something like a dashboard or task board. Look for the Interactive badge in the directory. Amplitude, Asana, Box, Canva, Clay, Figma, Hex and Slack have one today.

They're built on MCP Apps, an open extension to MCP. They run in sandboxed frames with the same permissions as the connector, and Anthropic doesn't allow purchases through them.

When one connector per app stops working

Directory connectors are the right start when the app you need has one. The app maker or Anthropic maintains them, and the sign-in is built for Claude.

They get awkward in a few situations we keep running into:

• You need an app that has no connector, and you're on Free, where one custom connector is all you get.
• You want Claude to search across apps at once, such as finding a decision that might sit in Slack, email or a GitHub issue.
• You want something to keep running after the chat ends, such as a summary posted to Slack every Monday.
• You want one approval rule for every app, instead of setting tool permissions connector by connector.

That's what we built Corsair for. Corsair is a remote MCP server at https://mcp.corsair.dev/mcp. You add it once as a custom connector, sign in to Corsair, and Claude gets the apps you've connected there, such as Gmail, Slack, GitHub, Linear, Notion, HubSpot and about 400 more. Each app has its own page listing the operations Claude gets, for example the Gmail MCP server.

Every Corsair connection uses a link with a mode, and Corsair checks each action against that mode, whatever the client allows:

• Read. Reads only. Writes are refused.
• Cautious. The default. Writes go through, and anything destructive, such as a delete, waits for your approval.
• Strict. Every write needs approval, and destructive actions are refused.
• Open. Everything runs without asking.

This sits on top of Claude's own tool permissions. You can keep Claude's write tools on Needs approval and still have Corsair refuse deletes outright.

Corsair can also set up automations that run on a schedule or when an app sends an event, and Claude can check how each run went. A directory connector only acts while you're chatting.

If the one app you need has a good directory connector, use it. Corsair earns its place when you need several apps, apps without a connector, or work that carries on after the chat. Setup for Claude and every other client is in Corsair's MCP docs.

If you're building a connector

We went through Anthropic's directory submission for Corsair this month. A few things the form checks aren't obvious from the MCP spec:

• Tool annotations. Every tool needs a title, plus readOnlyHint or destructiveHint. The form flagged a missing title on all 12 of our tools before we fixed it.
• A test account with real data. Reviewers sign in and use your tools. We set up a demo Slack workspace and a GitHub repo, and wrote prompts that exercise every tool.
• Sign-in that needs no setup. Claude can register itself through DCR or read its published identity through CIMD. Support either and people connect without entering a client ID.

Anthropic's guide to submitting to the Connectors Directory has the full list.

Troubleshooting

• The custom connector times out. Claude reaches the server from Anthropic's cloud, so a server behind a firewall or on a private network fails. Allowlist Anthropic's IP ranges, or use Anthropic's MCP tunnels.
• "Not found 404" while adding. The URL is wrong. Use the MCP endpoint itself, which often ends in /mcp.
• "Couldn't check the server". Claude couldn't work out how the server signs in. You can continue and set authentication by hand, but check the URL first.
• Sign-in fails after adding. Authentication settings can't be edited. Remove the connector and add it again with different options.
• Claude ignores the connector. Check that it's switched on for this chat under the + menu, and that its tools aren't Blocked.

FAQs

Does Claude Pro allow connectors?

Yes. Pro and Max can use directory connectors and add custom connectors by URL. The Free plan gets connectors too, with a limit of one custom connector.

What data can Claude access through connectors?

Only what the account you connected can access in that app, and only while the connector is switched on in your chat. A custom connector that uses one shared credential, such as an API key, reaches whatever that credential can reach.

What's the difference between a directory connector and a custom connector?

Directory connectors are listed in Claude's Connectors Directory and built by the app maker or Anthropic. A custom connector is any remote MCP server you add by its URL. Anthropic hasn't verified custom connectors, so only add servers from developers you trust.

Does Claude have connectors like ChatGPT?

Yes. Both are built on MCP, so one remote MCP server can work in both. The setup differs. In ChatGPT you add a custom MCP server, and on personal plans ChatGPT currently uses only a custom server's read tools.

Can I use Claude connectors in Claude Code?

Yes. Connectors work in Claude Code as well as in Claude on the web, Claude Desktop and the API. To add a remote MCP server from the terminal, run claude mcp add --transport http <name> <url>, then run /mcp in Claude Code to sign in.

How many connectors can I add?

The Free plan allows one custom connector. Anthropic's help center lists no cap for directory connectors, but it suggests switching tool access to On demand once 10 or more connectors are active.

Why won't my custom connector connect?

Claude connects to custom connectors from Anthropic's cloud, so the server must be reachable over the public internet. Check that the URL points at the MCP endpoint. If you picked the wrong sign-in options, remove the connector and add it again, because those settings can't be edited.