Claude MCP Server Guide 2026: Connect Claude to Your Tools

Last Updated: November 9, 2026

Summary: A Claude MCP server is a program that connects Claude to an outside tool or data source, such as a database, CRM or issue tracker, through the open Model Context Protocol. You can connect one in four places: as a custom connector in the Claude apps (available on every plan, with Free users limited to one), as a Claude Desktop extension, from the terminal in Claude Code with claude mcp add, or through the Claude API's MCP connector. Anthropic lists no separate charge for MCP; you pay for the Claude plan or API tokens you already use.

Searches for "Claude MCP server" have climbed steadily through 2026 as more teams try to give Claude live access to their own systems. Most of what ranks today is either developer documentation for one product or a list of favorite servers.

This guide is written for the business reader who has to decide how their team should use MCP with Claude. It puts the four connection methods side by side, shows which plans allow what and at what price, walks through setup, and covers the security and admin controls Anthropic documents. Everything was checked against Anthropic's own documentation on November 8, 2026.

A Claude MCP server is a small service that exposes tools and data to Claude using the Model Context Protocol, so Claude can do things like query a database, search an issue tracker or control a browser. The Claude Code documentation describes these servers as running "on your machine or as hosted services."

Anthropic introduced MCP on November 25, 2024, calling it "a new standard for connecting AI assistants to the systems where data lives." The launch announcement shipped a specification, SDKs, local server support in the Claude Desktop apps, and pre-built servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer.

MCP is no longer an Anthropic-only standard. On December 9, 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. At that point Anthropic counted more than 10,000 active public MCP servers and 97 million monthly SDK downloads across Python and TypeScript, and said ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code had adopted the protocol.

For a business, that matters because a server your team builds once can work with more than one AI product. If you are new to the agent side of this, our guide to what AI agents are explains why tool access is what turns a chatbot into an agent.

Claude can reach MCP servers through four routes: custom connectors in the Claude apps, Claude Desktop extensions, Claude Code, and the Claude API's MCP connector. Each one suits a different user, and they differ on whether the server runs locally or remotely.

The split between remote and local is the key decision. Anthropic's help article on custom connectors says remote servers must be reachable from Anthropic's IP ranges, and that connections come from Anthropic's cloud "across every Claude client, including claude.ai, Claude Desktop, Cowork, and the mobile apps." A local server, by contrast, runs on the user's own machine and never needs to be exposed to the internet.

Every Claude plan can use MCP through custom connectors, but the Free plan is capped at one custom connector and does not include Claude Code. Anthropic's help center says custom connectors are available "for users on Free, Pro, Max, Ultra, Team, and Enterprise plans."

Prices come from Anthropic's Claude pricing page, checked November 8, 2026, and the connector rules from Anthropic's custom connectors help article. The Team plan is listed "for teams of 2 to 150."

Here is what that means for a real budget. A 10-person team on Team Standard seats billed annually pays 10 x $20, or $200 a month ($2,400 a year). An Enterprise contract starts at the 20-seat minimum, or $400 a month in seat fees, before usage charged at API rates. Anthropic's pricing page lists no separate MCP charge on any of these plans.

On the API, there is no separate MCP fee listed either. Anthropic's MCP connector documentation has one pricing note: "MCP tool calls through the Batches API are priced the same as those in regular Messages API requests." Our Claude API pricing guide lists the current per-token rates.

In Claude Code, you add an MCP server with one terminal command, claude mcp add, then confirm it with claude mcp list. The steps below follow Anthropic's Connect to MCP servers guide.

Where a server is saved decides who gets it. Claude Code supports three scopes:

The project scope is the one teams care about. Committing .mcp.json to version control turns server setup into shared configuration, and Claude Code asks each teammate to approve a project server the first time it appears, so a cloned repository cannot launch processes on someone's machine without consent. Connectors added at claude.ai also load automatically in the Claude Code CLI when you sign in with the same account.

One cost to watch: Anthropic notes that each connected server takes space in Claude's context window, because its tool names and instructions load into every session. Remove servers you no longer use. For more on the tool itself, see our explainers on what Claude Code is and Claude Code plugins, which can bundle MCP servers for easier installs.

In the Claude apps, Pro and Max users add a remote MCP server under Customize > Connectors, while Team and Enterprise admins add it for the organization under Organization settings > Connectors. Anthropic says users "typically go through an OAuth authentication process" to grant access without sharing a password; servers that use API keys can receive fixed credentials through request headers instead.

For tools that live on a computer rather than in the cloud, Claude Desktop uses extensions. Anthropic's guide to local MCP servers on Claude Desktop describes extensions as "single-click installable packages" in the .mcpb format. You find them under Settings > Desktop app, then Manage next to Extensions, where "Browse extensions" opens a directory of Anthropic-reviewed extensions. Owners on Team and Enterprise plans can enable or disable public extensions and upload custom ones for their team.

Anthropic also runs a directory of reviewed connectors. The Claude Code documentation says directory connectors "use the same MCP infrastructure as Claude Code," so any remote server listed there can also be added from the terminal. Our guide to using Claude covers the rest of the app's features.

Developers can connect Claude to remote MCP servers straight from the Messages API, without running their own MCP client, using Anthropic's MCP connector. According to the MCP connector documentation, the feature is in beta on every model and is turned on with a beta header. On the Claude API, Anthropic now points developers to mcp-client-2026-09-15, which includes everything in the earlier mcp-client-2025-11-20 header and adds the option to pin each server's tool list. The 2025-11-20 header is still the one listed for Claude Platform on AWS, Claude Models on Google Cloud and Microsoft Foundry.

A request lists each server in an mcp_servers array (a type of "url", an https URL, a name and an optional OAuth token) and pairs it with an mcp_toolset entry in the tools array. The toolset can switch individual tools on or off, which lets you build an allowlist of read-only tools. Anthropic recommends denylisting write or destructive tools "when building read-only assistants, or when you want a human confirmation step before state changes."

Four limits matter before you build on it:

If you are building an agent rather than a chat feature, our step-by-step guide on how to build AI agents shows where tools like MCP fit in the design.

Claude MCP servers can be safe for business use, but only when you vet each server and limit what it can do, because Anthropic does not security-audit them. Anthropic's admin documentation for Claude Code says Anthropic reviews connectors against its listing criteria before adding them to the directory, "but doesn't security-audit or manage any MCP server."

The main risk is prompt injection. Anthropic's help center warns that custom connectors connect Claude "to arbitrary services that have not been verified by Anthropic," and that a malicious server may hide instructions that trigger unintended actions. The Claude Code documentation adds that servers which fetch outside content "can expose you to prompt injection risk." Based on Anthropic's own guidance, a basic checklist looks like this:

Admins have real controls. In Claude Code, an organization can deploy a managed-mcp.json file to give every user a fixed set of servers, publish an approved catalog with allowedMcpServers, block known-bad servers with deniedMcpServers, or disable MCP entirely with an empty server list. Anthropic warns that a server's name is "not a security control," so allowlists should match on server URLs or exact commands. On the claude.ai side, the pricing page lists admin controls for remote and local connectors on Team and Enterprise.

Sameer's take: In my conversations with public-sector organizations that are actively piloting AI, their biggest worry is privacy and data security. In my view, the biggest business mistake with AI is trusting output without checking it.

For a wider view of the policy side, see our guides to the risks of using AI at work and AI privacy.

Most business teams should start with custom connectors set up by a Team or Enterprise admin, and add Claude Code or the API only where developers need local tools or custom apps. The table below matches common situations to the route Anthropic's documentation supports best.

Is MCP free to use with Claude?
Yes. The Model Context Protocol is an open standard, and Anthropic says custom connectors work on the Free, Pro, Max, Ultra, Team and Enterprise plans. You pay only for the Claude plan or API tokens you already use.

Can Claude Free users add MCP servers?
Yes, with a limit. Anthropic says Free users are limited to one custom connector using remote MCP, and Claude Code is not included in the Free plan.

What is the difference between a Claude connector and an MCP server?
An MCP server is the service that exposes tools and data; a connector is how Claude's apps link to it. Custom connectors reach remote MCP servers over the internet, while Claude Desktop extensions and Claude Code can also run local MCP servers on your computer.

How do I add an MCP server to Claude Code?
Run claude mcp add in your terminal with a name and the server's URL, for example claude mcp add --transport http sentry https://mcp.sentry.dev/mcp. Check it with claude mcp list, then use /mcp inside a session to sign in if the server needs OAuth.

Who owns the Model Context Protocol now?
Anthropic created MCP in November 2024 and donated it in December 2025 to the Agentic AI Foundation. The foundation is a directed fund under the Linux Foundation, co-founded by Anthropic, Block and OpenAI.

A Claude MCP server is the bridge between Claude and your company's tools, and MCP is now an open standard that ChatGPT, Gemini, Cursor and Microsoft Copilot also support. Every Claude plan can connect at least one server, with the real choice being where: custom connectors for most staff, Desktop extensions for local files and apps, Claude Code for developers, and the API connector for your own products.

Anthropic lists no separate fee for MCP, so the budget question is seats and tokens. The harder work is governance: pick servers from vendors you trust, switch off tools you do not need, and use Team or Enterprise admin controls before rolling MCP out widely.

Written by Sameer Khan with AI assistance, fact-checked against the sources linked above.