Claude connectors: what they are and which to add

In claude.ai or Claude Desktop, go to Customize > Connectors, click the plus, browse the directory, open a connector and click Connect. You sign in to the other service, then switch the connector on per chat from the plus menu.

Anthropic's help article lists the same flow from inside a chat: click "+" or type "/", hover Connectors, pick Manage connectors. Web connectors work on Claude, Cowork, Claude Desktop and the mobile apps, and services you connect on web or desktop carry over to mobile when you next log in. Installing connectors on mobile is still in beta. On Team and Enterprise, an Owner has to enable a connector for the organisation first, and then each person signs in individually.

Claude Code is different. Per its docs, if you're logged in with a claude.ai account, the connectors you added at claude.ai are available automatically; run /mcp to see them. That doesn't happen if you authenticate with an API key or a cloud provider. You can also add a remote server directly:

claude mcp add --transport http <name> <url>

Two catches from the same page. A server you add locally takes precedence over a connector pointing at the same endpoint. And some Anthropic-hosted services (Microsoft 365, Gmail, Google Calendar) can't do local OAuth from Claude Code, so you connect those at claude.ai instead.

Directory connectors are listed by Anthropic and show up in the browse list; custom connectors are any remote MCP server URL you type in. Anthropic doesn't verify the custom ones.

A precise note, because page one of Google is vague on this. The directory doesn't label anything "official" or "community". The listings we read carry an "Anthropic verified" badge and a "Made by" line naming the vendor. That's as far as the labelling goes, and every listing also warns that Anthropic doesn't control what a developer's tools do or whether they'll change. So in this guide "directory" means listed and badged, and "custom" means you pasted a URL.

Adding a custom connector takes a name, the server's URL, and an authentication choice (sign in now, sign in when needed, or none). On Pro and Max it's Customize > Connectors > Add custom connector. On Team and Enterprise an Owner adds it under Organization settings first. Two things matter more than the clicks:

• It has to be public. Anthropic's cloud makes the connection, not your laptop, so a server behind a firewall or VPN won't connect unless you allowlist Anthropic's IP ranges. Local servers set up through claude_desktop_config.json aren't available in claude.ai or Cowork.
• Fixed keys are allowed. For servers that don't use OAuth, there's a request headers field for an API key or bearer token, sent on every request and stored encrypted.

A connector sees and changes only what you could, in that other app. Anthropic's help article says Claude inherits each person's permissions from the connected service. Reading and acting are separate: a connector can search your Drive or create pages in Notion depending on which tools it exposes.

A few details worth knowing before you connect anything that matters:

• Shared credentials are shared access. A custom connector that uses one API key gives everyone on the team that key's reach.
• Admins can narrow, not widen. On Team and Enterprise, Owners can set each connector's tools to Always allow, Needs approval or Blocked, grouped by read-only and write/delete. The person still needs permission in the source app.
• Research can run connector tools without asking. Anthropic's custom-connector article advises disabling tools that write to external apps before you use Research.
• Tool access has a mode. Auto is the default; Anthropic suggests On demand if you have ten or more active connectors.
• Prompt injection is real. Anthropic's warning is explicit: a malicious server can embed hidden instructions. Check tool inputs and outputs, and only choose Allow always for tools you'd trust unsupervised.

Our rule for marketing data: start every connector read-only. Let Claude draft the campaign change, then you make it.

Per Anthropic's help article, directory (web) connectors are available to all users on Claude, Cowork, Claude Desktop and mobile. Custom connectors are available on Free, Pro, Max, Team and Enterprise, with Free limited to one custom connector.

That's the stated position on 23 October 2026. Two caveats. The directory listings we read don't say which Claude plans they need, and the connected service may have its own account or plan requirements. And the Claude Code docs note that the claude.ai sync only works with a claude.ai login. If a limit matters for a purchase, check the help page that day.

A connector is an MCP server you reach through Claude's apps. A skill is written instructions. A plugin is a bundle that can include skills, agents, hooks and MCP servers.

Sources: Claude Code's MCP and plugins docs. Practical upshot: if you want Claude to reach your data, you need a connector or MCP server. If you want it to write like you or follow your process, you need a skill. A plugin is the way to ship both together. Our Claude for marketing guide covers where each Claude surface fits, and it's the place to start if you're not sure whether you want Projects, Cowork or Claude Code.

Pick by job, not by how many exist. The directory page we read showed 24 of 1,006 listings, so this isn't a ranking of everything. These are the ones we checked on their individual directory pages, each carrying "Anthropic verified" and a "Made by" line naming the vendor. We haven't run any of them.

How we'd use them. Ahrefs or Semrush answer the question Claude can't on its own: does anyone search for this? Our own SEO loop uses a keyword API for that reason, and a connector is the lighter version for people who already pay for one of those tools. HubSpot is the one to be careful with, since its tool names include creating and managing records and marketing email; keep those on approval. Adspirer is a good example of a sensible design, because it starts campaigns paused. Notion and Drive are where your brand notes and briefs probably already live, so Claude can read them without a copy-paste.

Some honest gaps. Figma's listing is aimed at generating code from designs, so it suits a developer more than a marketer. We found no social scheduling connector in the directory (Buffer and Hootsuite pages returned nothing when we looked, which isn't proof they don't exist elsewhere). We also found no Google Analytics listing at the address we tried. Bird, a messaging connector that can send email, SMS and WhatsApp, is listed, but sending bulk messages through a chat is a job for a human check first.

Disconnect and reconnect it first. Anthropic's checklist: check your internet connection and your account with the service, look at that service's permission or account-type requirements, then disconnect and reconnect.

If it's a custom connector and it times out, the usual cause is that the server isn't reachable from the public internet. To remove one, go to Customize > Connectors and click Remove (or the three dots). To change a custom connector's details, remove it and add it again.

If you use Claude Code and a connector shows as hidden, a locally added server with the same endpoint is taking precedence. Remove the local one, or leave it.

If you'd like to see the video step before you read on, you can try TopPost free and watch an edit come together.

TopPost is documented for Claude Code and other API-key clients. It may also work as a custom connector in claude.ai, but we haven't tested that, so we're not claiming it.

Here's what we can say. TopPost is a hosted short-form video pipeline: it edits your footage, A/B tests it as trial reels, schedules the winner and captures leads with auto DMs. Its MCP has 45 tools and API-key sign-in. The clients we document are Claude Code, Codex, Cursor, VS Code Copilot and Gemini CLI. Sign-in through claude.ai or Claude Desktop (OAuth) isn't live. Anthropic's article says a custom connector can send a fixed API key or bearer token as a request header, which is how TopPost authenticates, so a claude.ai custom connector could plausibly work. Plausibly isn't tested. If you try it, tell us what happened.

The Claude Code command:

Connect TopPost to Claude Code (create an API key first):

claude mcp add --transport http toppost https://app.toppost.ai/api/mcp --header "Authorization: Bearer <your API key>"

Any other MCP client: endpoint https://app.toppost.ai/api/mcp (streamable HTTP), header Authorization: Bearer <your API key>.

The limits you should know about:

• Your footage. It edits clips you upload; it doesn't generate video or avatars.
• Output. Vertical 9:16 MP4 only.
• Cost. Edits made by your own AI use no TopPost credits; you pay for storage. See pricing.
• Alternatives. The OpusClip connector above is a directory option for talking-head video. We haven't compared the two. How TopPost works explains ours end to end.

Everything in the marketing table. Reading a listing tells you what a vendor says the connector does; it doesn't tell you how it behaves with real data. A hands-on pass would check how often tools ask for approval, how HubSpot and Ahrefs handle a real account, what breaks on a Free plan, and what it costs in context when you've got ten connectors on.

Until then, treat this as a map, not a verdict. Pick one job, connect one read-only tool, and see if it earns its place before adding a second.