Claude Connectors: How to Connect Claude to Your Apps in 2026

Claude on its own is a very good writer that cannot touch a single thing you own. It can draft the caption but not post it, describe the design but not make it, summarise the spreadsheet only if you paste the spreadsheet in first.

Connectors are what close that gap. And the setup is genuinely a two-minute job, which is why it's frustrating that most guides either stop at "go to Settings" or disappear into developer documentation about JSON config files you'll never open.

This covers the whole thing: what a connector actually is, how to add one from the directory, how to add a custom one that isn't in the directory (the part nobody explains), whether they're safe, what they cost, and a worked example that ends with a real deliverable rather than a screenshot of a settings page.

What Is a Claude Connector?

A Claude connector is a permission you grant that lets Claude use one of your apps directly, so it can pull real data out or take real actions inside that app instead of just talking about it. Connectors run on the Model Context Protocol, an open standard Anthropic published so any tool can make itself usable by an AI assistant.

The mental model that works best: a connector is like creating a limited staff login inside a tool you already pay for and handing that login to an assistant. You decide which buttons it can press. It can't do anything you haven't allowed, it can't see anything your own account can't see, and you can take the login away at any time.

That framing matters more than the protocol details, because it tells you the two things you actually need to know before connecting anything: what can it reach, and how do I revoke it.

How Connectors Work

Three things happen when you connect an app, and understanding them makes the troubleshooting section later make sense.

Claude discovers what the app can do. On connection, the app tells Claude which actions it exposes. A design tool might offer "create a design," "edit a design," "export a file." A project tool might offer "list issues," "create issue." Claude reads that menu and now knows what's possible.

You authorise with OAuth, not a password. You get sent to the app's own login screen, you approve, and Claude receives a scoped token. Claude never sees your password. This is the same mechanism behind every "Sign in with Google" button you've ever clicked.

Permissions are inherited, not expanded. A connector cannot give Claude access to anything your own account doesn't already have. If you can't see a private project, neither can Claude.

How to Set Up a Claude Connector

There are two paths, and which one you need depends entirely on whether the app you want is already in Anthropic's directory.

Step 1: Open your connector settings

In Claude, go to Settings, then Connectors. You'll see anything already connected, plus a browsable directory of available apps. On Team and Enterprise plans, an admin has to enable connectors for the organisation before this section does anything, so if it looks empty or locked that's usually why.

Step 2: Connect a directory app

For anything in the directory, this is a two-click job. Find the app, click Connect, and you'll be redirected to that service's sign-in page. Approve the permissions it asks for and you'll land back in Claude with the connector live.

Test it immediately with something small and read-only. "What are my next three calendar events" is a better first test than asking it to reorganise your whole workspace.

Step 3: Add a custom connector by URL

This is the step almost every guide skips, and it's the one that matters most, because the directory is a fraction of what's available. Plenty of tools now ship their own MCP server and simply hand you a URL.

Click Add custom connector, give it a name, and paste the server URL. Then click Connect and approve on the other side.

Using Krumzi as a concrete example, the URL is https://www.krumzi.com/api/mcp. Name it Krumzi, paste, connect, approve. There's no API key to generate, no configuration file to edit, and no terminal involved. Custom connectors are available on Claude's Pro, Max, Team, and Enterprise plans.

One thing that trips people up: the URL usually needs the full path, not just the domain. Pasting the marketing site address instead of the server address is the single most common reason a custom connector fails to initialise.

Step 4: Check what it can actually do

Once connected, ask Claude directly: "what can you do with the Krumzi connector?" It will list the actions the server exposed. This is worth thirty seconds because it tells you the real boundaries rather than the ones you assumed, and it usually surfaces a capability you didn't know was there.

Step 5: Give it a real job

The test that tells you whether a connector is worth keeping is not whether it connected. It's whether it saves you a task you actually do.

For a design connector, that looks like:

Make me an Instagram post announcing our summer sale, 30% off, in our brand colours. Turn this article into a six-slide LinkedIn carousel. Design a two-page A4 menu for a coffee shop, print-ready.

A minute or two later the finished design is in your library, on brand, with every element still editable. That's the difference between a connector that's a novelty and one you keep.

Are Claude Connectors Safe?

Mostly yes, with one real caveat that deserves more attention than it usually gets.

The architecture is sound. OAuth means no password sharing. Scopes mean a connector can only do the specific things it declared. Inherited permissions mean it can't reach past your own access. And revocation is genuinely one click, in Settings under your connected apps.

The caveat is trust in the other party. A custom connector is a third-party service that Anthropic has not vetted. When you connect one, you're extending trust to whoever runs that server, the same way you would installing any other integration. Anthropic's own guidance is direct about this: don't connect a custom server unless you know and trust the application behind it.

There's also prompt injection to be aware of. If a connector pulls in content from an outside source, that content can in principle contain instructions aimed at the model. Reputable tools design around this, but it's a genuine category of risk rather than a theoretical one, and it's a good reason to be selective rather than connecting everything you can find.

Practical rules that cover most of it: connect tools you already pay for and already trust, start with read-only actions, check what a connector exposes before giving it real work, and disconnect anything you stopped using.

Are Claude Connectors Free?

There are two separate bills and conflating them causes most of the confusion.

On the Claude side, connectors need a paid plan. Custom connectors specifically are a Pro, Max, Team, and Enterprise feature. There's no per-connector charge on top.

On the app side, you pay whatever that app costs. A connector is a doorway into a tool, not a replacement for having an account. If the tool meters usage, work you trigger through Claude draws on the same meter as work you do in the app directly. Nothing is free that wasn't already free.

When a Connector Is the Wrong Answer

Worth saying plainly, because the enthusiasm around connectors tends to skip it.

If you only need the task done once, connecting an app is slower than just doing it. If the app has no MCP server, no amount of prompting creates one. If you need pixel-level control over a final asset, you'll be finishing in the app regardless, and the connector only saves you the first draft. And if the work is genuinely high-stakes and irreversible, keep a human between the request and the action.

Connectors pay off on repeated, low-stakes, well-defined tasks. That's a large category, but it isn't every category.

Frequently Asked Questions

Can Claude chat use connectors?

Yes. Connectors work in the regular Claude chat interface on the web and desktop apps, not just in developer tooling. Once you've connected an app in Settings, Claude can use it in any normal conversation. You don't need to do anything special to invoke it, you just describe what you want and Claude decides whether a connected tool can help.

What do connectors do in Claude?

They let Claude read from and act inside your other apps. Depending on what the app exposes, that can mean searching your files, querying data, creating or updating records, or producing an asset like a design or a document. Without a connector Claude can only work with what you paste into the conversation, which is the practical limitation connectors remove.

What are the best connectors for Claude?

The useful answer depends on what you repeat. For most people the highest-value connectors are the ones sitting on a daily task: a document or notes tool for anything writing-related, a project tool if you live in tickets, an analytics tool if you pull the same reports weekly, and a design tool if you produce visual content regularly. The worst connectors are the ones you set up because they were available rather than because you had a job for them.

Are Claude connectors safe?

The mechanism is safe: OAuth rather than passwords, scoped permissions, inherited access limits, and one-click revocation. The variable is who you connect to. Directory connectors come from established services, while custom connectors are third-party servers Anthropic hasn't reviewed, so the trust decision is yours. Connect tools you already use and pay for, and disconnect anything you've stopped using.

Can I connect an app that isn't in the Claude directory?

Yes, through Add custom connector. If the tool publishes an MCP server URL, you paste that URL, authorise, and it works exactly like a directory connector. This is how most newer tools ship support, since it doesn't require them to wait for a directory listing. You'll need a paid Claude plan for custom connectors.

The Takeaway

Connectors are the difference between an assistant that describes your work and one that does it. The setup is two minutes, the security model is sound as long as you're deliberate about who you trust, and the real cost is on the app side rather than Claude's.

Start with one connector attached to something you do every week, not five attached to things you might do someday. If design is that weekly task, the Krumzi MCP connection takes about two minutes and ends with a finished, editable design rather than a description of one. For more on why that layered output matters, our guide to agentic AI design covers the difference between a tool that generates pictures and one that generates work.

Anthropic's documentation on using connectors to extend Claude's capabilities is worth a read for the platform-side specifics, and the Model Context Protocol announcement explains the standard underneath if you want the background.