Claude MCP: how to connect your apps without editing a config file

MCP is the open standard Claude uses to reach your other software, and inside Claude’s own interface it is labeled “connectors” rather than MCP, which is why searching your settings for the letters M, C, P finds nothing. Turning one on is a settings task rather than a coding task: on claude.ai you open Customize, then Connectors, pick a service, click Connect, and sign in to that service the way you normally would (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22). A connector cannot reach anything your own account in that app cannot already reach, because Claude inherits your permissions from the connected service, so the question worth asking before you click is not what Claude will be able to see but what you want it to be able to change (same source, retrieved 2026-08-22).

What that buys you is the part the search results do not say. Claude reaches Notion, Slack and Google Drive today without Notion, Slack or Google having shipped an AI feature for Claude. The reach is a property of an open standard the applications publish to, not a feature any of those vendors decided to build for you, which is why it does not wait on their roadmaps. The longer version of that argument is on the manifesto.

This page is about the Claude you talk to: claude.ai in a browser, the Claude desktop app, and the mobile apps. It is not about Claude Code, the command-line tool for programmers. If you have never opened a terminal, you are in the right place and you will not need one.

What is MCP, and why does Claude call it something else?

MCP stands for Model Context Protocol, an open standard for connecting AI applications to outside systems, described by its own documentation as a standardized port for AI applications in the way USB-C is a standardized port for devices (modelcontextprotocol.io, retrieved 2026-08-22). Anthropic published it on 2024-11-25 and made it open (anthropic.com, 2024-11-25), then donated it in December 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation, where the existing maintainers keep full autonomy over technical direction (blog.modelcontextprotocol.io, retrieved 2026-08-22). So this is not Anthropic’s thing to withdraw or reprice, and it is now supported by other assistants and editors including ChatGPT, Visual Studio Code, and Cursor (modelcontextprotocol.io, retrieved 2026-08-22).

The naming gap is the practical problem. MCP is the protocol, the plumbing underneath. What you see in Claude is a connector, which Anthropic’s documentation describes as being powered by MCP (Anthropic, Connectors overview, retrieved 2026-08-22). So when a tutorial says “add an MCP server to Claude” and your Claude has no such button, nothing is broken and no feature is missing from your plan. You are looking for the word connector.

Two other words show up in the same neighborhood and mean different things. A desktop extension is a connector that runs on your own computer rather than in the cloud, for reaching local files or an application installed on your machine (Anthropic, When to use desktop and web connectors, retrieved 2026-08-22). A plugin is a package that can bundle connectors together with other material, and plugins are available in Claude Code and Claude Cowork (Anthropic, Connectors overview, retrieved 2026-08-22).

Where do I find MCP in the Claude I already use?

In two places, and both are in the ordinary settings interface. From a chat, click the + button in the lower left, or type /, hover over Connectors, and select Manage connectors. From settings, go to Customize > Connectors at claude.ai/customize/connectors. Either route opens the same catalog, which Anthropic calls the Connectors Directory, browsable by category or as a full list (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22).

Anthropic’s documentation states that web connectors are available for all users on Claude, Claude Cowork, Claude Desktop, and Claude Mobile on iOS and Android, and that desktop extensions are available to all users on Claude Desktop (same source, retrieved 2026-08-22). A second Anthropic page states a different prerequisite, and that disagreement is unresolved as of the date on this page. See the plan question below.

How do I turn on my first connector?

Five clicks and one sign-in, with no file to edit. Anthropic documents the sequence as: click the connector you want, review its description and capabilities, click Connect or Install, follow the authentication prompts to grant Claude access to your account, then configure any specific settings or permissions (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22). The authentication step is the standard sign-in screen for the app you are connecting, the same one you would see when a phone app asks to use your Google account.

Once it is connected, you turn it on for a given conversation. Click the + in the lower left of the chat or type /, hover over Connectors, and toggle on the services you want Claude to use for that conversation (same source, retrieved 2026-08-22). Connecting and enabling are separate acts, which is useful: a connected service is available, not automatically in play.

Two things worth knowing before you start. If you are on a Team or Enterprise plan, an Owner or Primary Owner has to enable a connector for the organization before you can authenticate to it, and enabling it for the team does not grant anyone access, since each person still authenticates individually (same source, retrieved 2026-08-22). That is the likeliest reason a connector you read about is not in your list, and the ask is specific enough to put in a message to whoever administers your account. And connections expire: the documented fix is to go to Customize > Connectors and click Reconnect or Refresh (Anthropic, Connectors getting started, retrieved 2026-08-22).

Prebuilt first-party connectors documented by Anthropic include Google Drive, Gmail, Google Calendar, GitHub, Slack, and Microsoft 365, described as ready to use with no setup beyond authentication (Anthropic, Connectors overview, retrieved 2026-08-22). For what each of those applications actually exposes once connected, the per-application pages carry the action surface and the date it was checked: Slack, GitHub, and the rest through the tools index.

Is it safe to connect my email or my files?

The single most useful fact is that a connector does not widen your access. Anthropic’s documentation states that Claude inherits each person’s permissions from the connected service, and that if someone cannot access a specific file, channel, or record in the source system, the connector cannot reach it from Claude either (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22). Connecting Google Drive does not hand Claude the whole company’s Drive. It hands Claude your view of it.

What does change is that something other than you can now act with your access. Anthropic’s own warning is direct: when you connect to a service you are granting Claude permission to access and potentially modify data within that service based on your account permissions, and you should only connect services you trust and need (same source, retrieved 2026-08-22). Read that as a scope question rather than a yes or no question. Reading your calendar and sending mail from your address are different risks that arrive through the same door.

Who built the connector matters, and Claude labels it. Anthropic’s directory carries three states: Verified, meaning Anthropic tested the connector’s tools for quality and compatibility against its Software Directory Policy at the time of review, which the documentation explicitly says is not a security audit; Community, meaning a third-party developer built it and Anthropic screened but did not review it in depth, with a reminder shown before you connect; and Custom, meaning you added it yourself and Anthropic has not reviewed it (Anthropic, Connector verification, retrieved 2026-08-22). One line on that page is the one to keep: the label affects how a connector is displayed and discovered, not how it functions, and once connected a community connector has the same capabilities and access as any connector you grant (same source, retrieved 2026-08-22).

Anthropic’s advice for any connector built by someone other than Anthropic is worth reproducing plainly, because it is the operating instruction rather than a disclaimer. Connect only to developers and organizations you trust. The developer controls which tools the connector exposes and can change them at any time. Anthropic does not run a third-party connector’s servers and does not control how it handles your data. Review the requested permission scopes during authentication. Be aware of prompt injection risks, for which Claude has built-in protections. Watch for unexpected changes in tool behavior (same source, retrieved 2026-08-22).

On data handling, the documentation states that all data transfers are encrypted, that you can only sync content you have permission to view in the original source, and that connected services process data on their own infrastructure under their own terms, which may sit outside the United States. Settings that control where Claude’s inference runs, such as the US-only inference setting on Enterprise plans, do not change where a third-party service operates (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22).

Does this cost anything?

Two Anthropic pages give different answers, and both were retrieved on 2026-08-22. The support article states that web connectors are available for all users, and that custom connectors using remote MCP are available to users on free, Pro, Max, Team, and Enterprise plans, with free users limited to one custom connector (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22). The developer documentation’s setup tutorial lists as a prerequisite “A Claude account (Pro, Max, Team, or Enterprise for most connectors)” (Anthropic, Connectors getting started, retrieved 2026-08-22).

Those cannot both be a complete description of the same thing. We are reporting the conflict rather than picking the more convenient side, and the tutorial page carries other signs of being older than the support article, including a statement that Claude can search your email but cannot send it, which sits against the support article’s own example of a connector sending a Slack message. Check your own Customize > Connectors screen, which is authoritative for your account in a way that no third-party page can be.

Separately, MCP itself is an open standard and costs nothing to use. The protocol’s public registry is a paginated open API whose limit parameter caps at 100 results, and walking all 243 pages with version=latest on 2026-08-22 returned 24,221 distinct server names (registry.modelcontextprotocol.io/v0/servers, full pagination 2026-08-22). What can cost money is the app on the other end of the connection, since a connector reaches your existing account there and does not create one.

What can I ask for once a connector is on?

Ordinary requests in ordinary words, naming the thing you want rather than the tool. Anthropic’s documented examples are connecting Claude to Linear to create issues, to Slack to send messages, and to Google Drive to search your files, with the worked example being to ask Claude to “Create a new issue for the login bug” after connecting Linear (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22). You do not have to name the app every time either. Once a service is connected, Claude can bring it into a conversation on its own when it fits the request (same source, retrieved 2026-08-22).

Some connectors return an interface rather than text. Anthropic describes interactive connectors that render live dashboards, task boards, and design tools inside the conversation, marked with an Interactive badge in the directory (same source, retrieved 2026-08-22).

If you connect a lot of services, there is a setting that stops them crowding each other out. Under + > Connectors > Tool access you choose Auto, the default, where Claude decides which connectors to load based on what you are working on; Always available, which loads all of them at the start of every conversation and uses more conversation space upfront; or On demand, which loads nothing until Claude searches for the right connector, at the cost of one extra step. Anthropic recommends Auto for most people and On demand once you have ten or more connectors, and the setting applies to the single conversation you set it in (Anthropic, Manage Claude’s tool access, retrieved 2026-08-22).

Do I have to be good at the app before this helps?

No, and the assumption that you do is the most common reason people stop at the connector screen. AI capability does not sit on top of tool expertise. What Claude lacks is not skill with Google Drive or Slack, it is knowledge of your work, and four answers cover most of that. They are the same four used on every guide on this site.

1. What is this task’s one job?
2. What does a good result look like, specifically enough that someone else could tell?
3. What must never happen without my approval?
4. What supervision does this task need?

A connector that has been asked a half-specified question will ask which account or which record, and that is the mechanism working rather than failing.

How do I stop Claude from doing something I did not want?

Start with the two controls every account has: which connectors are toggled on in the conversation you are in, and whether the service is connected at all. Disconnecting, changing connection settings, and reviewing permissions and access levels are all documented under Customize > Connectors (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22). Turning a connector off for a conversation is not an admission of anything. It is the normal way to keep a task narrow.

Finer control over individual actions is documented for organizations. Anthropic describes Owners on Team and Enterprise plans limiting which actions a connected service can take across the organization, with permissions grouped by type such as read-only and write or delete, and each set to Always allow, Needs approval, or Blocked. The documented examples are allowing Claude to search and summarize email while preventing it from sending, and allowing it to read Google Drive files while preventing it from creating or editing documents. That setting applies organization-wide and individual users cannot override it (same source, retrieved 2026-08-22). Whether an individual on a personal plan sees the same per-action control is not stated in the documentation we read, and we are not asserting either way.

The floor underneath all of it is the permission inheritance above. Action restrictions in Claude work alongside the source system’s own permissions and can only narrow them, never widen them, so a write you allow in Claude still fails if your own account in that app cannot make the change (same source, retrieved 2026-08-22).

What if the app I need is not in the directory?

The directory is a catalog rather than a requirement. Anthropic’s documentation states that connectors in the directory and custom connectors you add yourself use the same technology, that any connector with a URL can be added as a custom connector, and that a connector does not need to be in the directory for you to use it (Anthropic, Connector verification, retrieved 2026-08-22).

Adding one by hand is still a settings task. Go to Customize > Connectors, click the + next to Connectors, select Add custom connector, enter the connector’s name and URL, optionally enter an OAuth client ID and secret under advanced settings, click Add, and complete the same connection flow as any directory connector (Anthropic, Use connectors to extend Claude’s capabilities, retrieved 2026-08-22). One constraint catches people out: custom connectors reach your server from Anthropic’s cloud rather than from your own device, even in Claude Desktop, so a server behind a corporate firewall or on a private network will fail to connect (same source, retrieved 2026-08-22).

This is the point where the tutorials you have already found start talking about a JSON file, and it is worth naming why. Editing claude_desktop_config.json is the route for a server running on your own machine, which is a real and sometimes necessary path, and it is also the route almost every page on this topic sends every reader down regardless of what they were trying to do. If the thing you want to reach is a cloud service you sign into, the remote connector above is the documented route and it works across web, mobile, desktop, and Claude Code (Anthropic, When to use desktop and web connectors, retrieved 2026-08-22).

Where a URL is the thing you do not have, a managed connector service is a third route, and the difference from the config-file route is where the credential lives rather than what the assistant can do. An API key pasted into a config file sits in a file on your machine, is as broad as whatever it was issued for, and expires without announcing it. A managed connector holds the authorization outside that file, so a break surfaces as a connection marked expired rather than as a task that quietly stopped working. Either way the rule is the same: no key, token, or session cookie belongs in a file you sync or share.

Two figures for scale, from a catalog inventory of software reachable that way, built 2026-08-14 and queried 2026-08-22: 589 published applications out of 1,000 in the catalog, with a median of 19 actions retrieved each. The median matters more than the count. Most applications expose enough actions for one bounded job and not enough for an open-ended one, which is a better guide to what to try first than any list of the most connectable apps.

This page stops at what one person can find out and ask for. Anything that promises a result across a whole team is a different kind of project with a different kind of budget, and it is not what a page like this can deliver.

Do I need Claude Code for any of this?

No, and this is the confusion the search results create. Claude Code is the command-line tool for programmers, and it is what the top-ranking pages on “claude mcp” are documenting. Remote connectors, the kind you add from the directory, work across every Claude surface: web, mobile, Claude Cowork, Claude Desktop, and Claude Code, with no extra setup once connected (Anthropic, When to use desktop and web connectors, retrieved 2026-08-22).

The one case that genuinely requires a local surface is a tool running on your own computer. Desktop extensions run locally and are only available in Claude Desktop and Claude Code, not on web or mobile (same source, retrieved 2026-08-22). Anthropic’s own guide for choosing between them:

Source: Anthropic, When to use desktop and web connectors, retrieved 2026-08-22.

What is a connector the wrong tool for?

Anything where you would not be able to tell whether the result was right. That is the honest boundary, and it does not move as you get more comfortable. A connector is very good at reaching something you already have permission to reach and reporting back. It is a poor fit for a judgment that depends on context you have never written down.

Four ways of working, each a permanent right answer for some tasks and a wrong answer for others:

None of these is a step toward the next one. Read-only is the correct permanent setting for plenty of work, and a fully specified procedure is the wrong answer for anything where the definition of a good result is still moving. The full version of the table, with what each shape is wrong for, is on the manifesto.

Three specific cases to leave alone for now. Work covered by a rule you did not write, where the constraint is a policy question rather than a settings question, and the useful move is asking whoever owns the policy rather than turning something on. Work where a mistake is not visible until much later, since the checkable part is what makes a bounded task safe. And a task you are only automating because it is annoying, which is often a sign that the task itself should be smaller. Simplifying it is a complete outcome, not a delay.

What could not be established here?

• Which plan is actually required. Two Anthropic pages retrieved on 2026-08-22 disagree, as described above: the support article says web connectors are available to all users and that free users get one custom connector, while the developer tutorial lists Pro, Max, Team, or Enterprise as a prerequisite for most connectors. We did not resolve it in favor of either, and no account-level test was run to break the tie.
• Whether an individual on a personal plan can set per-action permissions. Anthropic documents Always allow, Needs approval, and Blocked for Owners on Team and Enterprise plans, applied organization-wide. The documentation we read does not state what an individual on Free, Pro, or Max sees, and we did not test it.
• Which named connectors are the vendor’s own. Full pagination of the MCP registry returned 24,221 distinct server names on 2026-08-22, and registry presence does not establish that a server is first-party. Forty-one applications have been checked one at a time on the best MCP servers page; no check was run for this page, so it names no third-party connector as official.
• How many of the 589 a person can turn on unaided. The 589 published applications were not checked for how many are reachable without an administrator enabling something first. That count does not exist yet.
• Whether any of this saves time, and how much. No measurement was run. Any time saving attached to connectors, including any we might publish later, is an estimate until it is measured against your specific task.
• How well the prompt injection protections work. Anthropic states that Claude has built-in protections against prompt injection (Anthropic, Connector verification, retrieved 2026-08-22). We found no published effectiveness figure and ran no test, so treat that as a stated protection rather than a measured one.
• What the interface looks like today. Every click path here is transcribed from Anthropic’s documentation retrieved 2026-08-22, not from a screen recording made on that date. Consumer AI interfaces move faster than the pages that describe them, and if a menu name has changed, your own Customize > Connectors screen wins over this page.
• Demand for this question. claude mcp returns no measurable volume on Google Ads and 334 in the AI channel, against 64 twelve months prior, measured 2026-08-22, United States, English. Absence of reported Google volume is not absence of demand, since Google Ads omits keywords below a reporting threshold, and the AI channel figure is modeled from People Also Ask data rather than counted from assistant traffic.

Sources