Claude Administration
Claude MCP: what it is, how connectors use it and how to govern it
A plain guide for IT and operations leads: what the Model Context Protocol is, how a Claude connector relies on it, and what to settle before a team connects Claude to company systems. Taken from the MCP project's and Anthropic's own pages as read on 15 November 2026.
Claude MCP means using the Model Context Protocol (MCP) to connect Claude to other systems. The MCP project describes MCP as an open-source standard for connecting AI applications to external systems, and Anthropic's help article says Claude can connect to any service that supports MCP. In our reading, a Claude connector is how a person or an organization turns that on for one service. Before a team connects Claude to company systems, it should settle which servers are allowed, who may add them, what each one may change and who reviews the result. Sources read on 15 November 2026.
plain answer
What does MCP do for Claude?
MCP is a standard that lets an AI application such as Claude reach systems outside the conversation. The MCP project's introduction says MCP is "an open-source standard for connecting AI applications to external systems", and that AI applications like Claude can use it to reach data sources such as local files and databases, tools such as search engines, and workflows such as specialized prompts. The same page compares MCP to a USB-C port: one standard way to plug an AI application into many different systems.
The MCP architecture page names three parts. The host is the AI application, for example Claude Code or Claude Desktop. The host creates one client for each server it connects to. The server is the program that provides context. A server can offer three kinds of things: tools (functions the AI can run, such as an API call or a database query), resources (data the AI can read) and prompts (reusable templates).
Sources: MCP project, "What is the Model Context Protocol (MCP)?" (https://modelcontextprotocol.io/docs/getting-started/intro) and "Architecture overview" for protocol version 2026-07-28 (https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture), both read 15 November 2026.
plain answer
How does a Claude connector use MCP?
Anthropic's help article on connectors says connectors "let Claude access your apps and services, retrieve your data, and take actions within connected services." It also says you can connect Claude to any service that supports MCP, and that custom connectors use remote MCP servers. Anthropic's pages do not define a connector in those words. In our reading, a connector is how Claude is set up to use one service, and for a custom connector that means pointing Claude at the address of an MCP server.
In plain terms, when someone asks Claude a question that needs a connected service, Claude can call the tools that service's MCP server offers (MCP architecture page). What it can do is limited twice. The article says Claude inherits each person's permissions from the connected service, and that restricting actions in Claude never grants more access than the source system permits.
Claude Code, Anthropic's coding tool, works the same way. Its documentation says Claude Code can connect to external tools and data sources through MCP, and that MCP servers give it access to your tools, databases and APIs. The Anthropic connector directory is where many ready-made connectors are listed, and our page on Claude connectors walks through twelve of them. See also our page on Claude Code with your dev stack.
Sources: Anthropic Help Center, "Use connectors to extend Claude's capabilities" (https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities) and Claude Code documentation, "Connect Claude Code to tools via MCP" (https://code.claude.com/docs/en/mcp), both read 15 November 2026. The tool-call sentence is from the MCP architecture overview (URL above), read the same day.
plain answer
What is the difference between a local and a remote MCP server?
The MCP architecture page says a server is a program that serves context, wherever it runs. A local server runs on the same machine as the client and uses the stdio transport, so it typically serves one client. A remote server runs on another platform and uses the Streamable HTTP transport, so it typically serves many clients. The page's own examples are a filesystem server (local) and the Sentry server (remote).
The difference matters for risk. Anthropic's Claude Code documentation says stdio servers run as local processes on your machine, and recommends HTTP servers for connecting to remote services. The MCP security guidance warns that local servers may have direct access to the user's system, because they run with the same privileges as the client.
One detail from Anthropic's help article is easy to miss. For custom connectors, the connection to your MCP server is made from Anthropic's cloud, not from your device, and the server must be reachable over the public internet. For a server behind a firewall, the article says you can allowlist Anthropic's IP ranges in your firewall, and it links to another article for the network requirements. Check the article before you plan this, because plan details change.
Sources: MCP architecture overview (https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture), MCP security best practices (https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices), Claude Code MCP documentation and Anthropic Help Center connectors article (URLs above), all read 15 November 2026.
plain answer
Who can add a connector in a company Claude plan?
Anthropic's help article separates individuals from organizations. An individual can browse the directory and connect a service by following the sign-in prompts. On Team and Enterprise plans, an Owner or Primary Owner must first enable a connector for the organization. Enabling it does not automatically give anyone access, and each person still signs in to the service themselves. For the plans themselves, see our page on Claude Team vs Enterprise.
People who cannot add connectors can send a request. Anthropic's Help Center says (read 15 November 2026) that requests are on by default for Team plans and off by default for Enterprise plans, and that only Owners and Primary Owners can change that setting. Owners can also set a tool permission for the whole organization: Always allow, Needs approval or Blocked, which individual users cannot override. The article labels one further option, enterprise-managed authentication, as beta.
For Claude Code, the documentation describes three places a server can be defined: local (just you, in the current project), project (shared through a .mcp.json file in the project) and user (you, across all projects). It says Claude Code asks for approval before using project-scoped servers from that file in interactive sessions, but loads them without asking in non-interactive runs, and it describes a setting for blocking a named server. It also says administrators can provide servers to every user through managed configuration. We did not read the full allowlist and denylist mechanics, so check Anthropic's page for those.
Sources: Anthropic Help Center connectors article and Claude Code MCP documentation (URLs above), read 15 November 2026. For how Claude's admin area is organized, see our page on the Claude admin console.
plain answer
How should a team govern MCP use in Claude?
Neither the MCP project nor Anthropic gives a one-line rule, but their pages point to the same few decisions. This is our reading of them, not a standard.
Decide who may add a server. Anthropic's help article advises connecting only services you trust and need, and warns that custom connectors can reach services Anthropic has not verified. Claude Code's documentation says to verify you trust each server, because servers that fetch external content can expose you to prompt injection. For more on that term, see our page on prompt injection.
Give each connection the narrowest access that works. The MCP security guidance says to start with a minimal set of permissions and raise them only when a task needs it, and warns against broad omnibus scopes because a stolen token then reaches much more. Anthropic's tool permissions (Always allow, Needs approval, Blocked) are the place to apply this for Claude.
Treat local servers as software you are installing. The MCP guidance says that when a client offers one-click setup of a local server, it must show the exact command, without truncation, and get explicit approval first. It also says clients should run local servers in a sandbox with restricted access to files and the network.
Know where a server's tokens go. The MCP guidance says a server must not accept tokens that were not issued for it, which matters if you build or buy a server that sits between Claude and another system.
Keep a written list. Record each approved server, who owns it, what it can read and change, and who reviews its output. This is a practice we recommend; neither source requires it. For the detail on each risk, read our page on MCP security.
Sources: MCP security best practices (https://modelcontextprotocol.io/specification/2026-07-28/basic/security_best_practices), Anthropic Help Center connectors article and Claude Code MCP documentation (URLs above), read 15 November 2026.
plain answer
Where Kriv helps
Kriv AI helps teams decide which MCP servers and connectors to approve, set the permissions above, and write the list and the review steps down before Claude is rolled out widely. This page is not legal advice. Kriv AI has achieved the Select partner tier in the Claude Partner Network Services Track.
If you want to see how this looks for one real connector, our page on the Claude AWS integration walks through AWS MCP. For other systems, see Claude and Jira, Claude and Salesforce, and Claude and Notion. See also our pages on AI agent integration and Claude enterprise training.
limits
What this page could not confirm
We read the MCP project's documentation and Anthropic's help article and Claude Code documentation on 15 November 2026. None of the pages show a publication date, so we cannot say how recent they are beyond our read date. The MCP pages describe protocol version 2026-07-28.
We did not read Anthropic's page on building connectors, its managed-configuration page for Claude Code, or the MCP authorization specification, so this page does not describe how sign-in is set up or how allowlists work in detail. The connector directory page gave no definition of a connector, so we did not use it. We have not tested any connector, so we make no claim about how a particular one behaves. Plan features, defaults and beta labels change; check Anthropic's pages before you act.
Sources
Cited sources
- MCP project: What is the Model Context Protocol (MCP)? (no date shown), read 2026-11-15
- MCP project: Architecture overview, protocol version 2026-07-28 (no publication date shown), read 2026-11-15
- MCP project: Security Best Practices, version 2026-07-28 (no date shown), read 2026-11-15
- Anthropic Claude Code docs: Connect Claude Code to tools via MCP (no date shown; first 100,000 of 156,945 characters read), read 2026-11-15
- Anthropic Help Center: Use connectors to extend Claude's capabilities (shows 'Updated over a week ago', no date), read 2026-11-15
Straight answers
Frequently asked questions about Claude MCP: what it is, how connectors use it and how to govern it
What is MCP in Claude?
MCP is the Model Context Protocol, which the MCP project describes as an open-source standard for connecting AI applications to external systems. Claude uses it to connect to data sources, tools and workflows (MCP project introduction, read 15 November 2026).
Is a Claude connector the same as an MCP server?
In our reading, no: the MCP server is the program that provides tools and data for one service, and the connector is how Claude is set up to use it. Anthropic says Claude can connect to any service that supports MCP and that custom connectors use remote MCP servers (Anthropic Help Center, read 15 November 2026).
Can any employee add an MCP connector to a company Claude plan?
On Team and Enterprise plans, Anthropic says an Owner or Primary Owner must first enable a connector for the organization, and each person still signs in themselves. Defaults for requests differ by plan (Anthropic Help Center, read 15 November 2026).
Does a custom connector reach a server on our private network?
Anthropic's help article says custom connectors connect from Anthropic's cloud, not from your device, and the server must be reachable over the public internet. For a server behind a firewall, it says you can allowlist Anthropic's IP ranges. Check the article for current detail (read 15 November 2026).
Does MCP remove the risk of prompt injection?
No. Claude Code's documentation says servers that fetch external content can expose you to prompt injection risk, so it advises verifying that you trust each server (Claude Code documentation, read 15 November 2026).
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call