How MCP Lets an AI Assistant Use Your Business Apps
MCP is the open standard that lets an AI assistant read from and act in other apps. Its parts, consent rules and setup steps, with Stripe’s server as the example.
Checked against the MCP specification and docs, version 2026-07-28 (latest), and Stripe’s MCP documentation
As of · MCP specification version 2026-07-2823
| Piece | What it is | Who decides when it runs | Example in Stripe’s server | Source |
|---|---|---|---|---|
| Tools | Executable functions the AI application can invoke to perform actions, such as file operations, API calls or database queries | The language model can discover and invoke them from the conversation; the spec says a human should be able to deny each call | stripe_api_read, stripe_api_write and get_stripe_account_info | 2, 4, 7 |
| Resources | Data sources that give the AI application context, such as file contents, database records or API responses | Application-driven: the host application decides how to bring them in, for example by letting you pick them | Not listed on Stripe’s page | 2, 5, 7 |
| Prompts | Reusable templates that help structure an interaction, such as system prompts or few-shot examples | User-controlled: you explicitly select them, usually by a command in the interface | Not listed on Stripe’s page | 2, 6, 7 |
| Elicitation | The server asks the user for more input, or for confirmation of an action (a client feature) | The user answers; the server only asks | Stripe’s own approval link for actions such as refunds and outbound payments (a Stripe step) | 2, 7 |
The first three rows are the primitives a server can expose; elicitation is a primitive a client exposes. The protocol deprecates two other client primitives, sampling and logging, as of version 2026-07-28.2 “Who decides” is the interaction model each spec page describes, not a guarantee of how any given app behaves: the spec says implementations may expose tools through any interface and that MCP itself can’t enforce its consent principles.43 Stripe’s list of tools is its own, includes preview features and may change.7 This post describes the protocol; it isn’t security or legal advice.
How it works
A shared way to plug an AI into other apps
MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems: data sources such as local files and databases, tools such as search engines and calculators, and workflows such as specialized prompts.1 The project’s own comparison is a USB-C port: one standard way to connect, instead of a custom link between every AI app and every service.1 It covers only the protocol for exchanging context; it does not dictate how an AI application uses a language model or manages the context it receives.2
Who is who: host, client, server
An MCP host is the AI application, such as Claude Desktop or Visual Studio Code. It creates one MCP client for each MCP server it connects to, and the server is the program that provides context.2 A server can run on your own machine (a “local” server, using standard input and output) or on the app maker’s platform (a “remote” server, using Streamable HTTP); the docs’ examples are a local filesystem server and Sentry’s remote server.2 Remote servers can serve many clients and are available from any MCP client with an internet connection, without installing anything on each device.8
What a server can offer
Servers expose three primitives, listed in the table: tools, resources and prompts. A client first asks which are available and then uses them, so the lists can change; a database server, for instance, might offer tools to query it, a resource holding its schema and a prompt with examples.2 A server can also ask the user for more input or confirmation through elicitation.2
Where you stay in control
The specification’s key principles say users must explicitly consent to and understand all data access and operations, hosts must obtain explicit consent before exposing user data to servers, and hosts must obtain explicit user consent before invoking any tool.3 It treats tools as arbitrary code execution, so descriptions of what a tool does should be considered untrusted unless they come from a trusted server.3 For tools the spec says there SHOULD always be a human in the loop who can deny an invocation, with clear indicators when a tool is used and confirmation prompts for operations.4 The spec admits MCP can’t enforce these at the protocol level; they fall to each application.3
Signing in, using Stripe as the example
Remote servers usually need authentication, commonly OAuth, API keys or a password, and the MCP docs say to verify a server’s authenticity and review the permissions it asks for.8 Stripe’s server at mcp.stripe.com supports OAuth, for a client acting for your Stripe user account, and agent API keys, for an AI application that needs independent, persistent access.7 With OAuth, Stripe’s consent page lets you grant access to one or more live accounts or sandboxes with different permissions for each, you can list and revoke authorized clients under OAuth sessions, and administrators can switch MCP access on or off for a whole team.7 Stripe also requires human confirmation before certain write actions, such as refunds and outbound payments, through a link the agent gives you, which expires after 24 hours.7 Stripe says that from December 2026 its MCP server stops accepting full-access secret keys and restricted keys without the Agent tag, so those need replacing with agent API keys or OAuth by November 30, 2026.7
The setup effort, as documented
The effort is a few steps in the AI application, not code. MCP’s guide for connecting Claude to a remote server lists five: open the connector settings, add a custom connector with the server’s URL, complete authentication, then use its resources and prompts and set tool permissions.8 For Stripe in Claude Desktop, Stripe lists three: connect the connector, sign in and grant permissions through OAuth, then enable Stripe in a conversation.7 Stripe also offers a one-command route, stripe agent setup, which detects the agents you use and configures them.7 Stripe’s page tells users to turn on their client’s tool-approval setting to review calls before they run, and to be careful when using its server alongside other servers because of prompt-injection attacks.7
Sources
- Model Context Protocol, What is the Model Context Protocol (MCP)?, version 2026-07-28, retrieved 2026-10-28.
- Model Context Protocol, Architecture overview, version 2026-07-28, retrieved 2026-10-28.
- Model Context Protocol, Specification overview, including “Security and Trust & Safety” key principles, version 2026-07-28, retrieved 2026-10-28.
- Model Context Protocol, Specification: Tools (user interaction model), version 2026-07-28, retrieved 2026-10-28.
- Model Context Protocol, Specification: Resources, version 2026-07-28, retrieved 2026-10-28.
- Model Context Protocol, Specification: Prompts, version 2026-07-28, retrieved 2026-10-28.
- Stripe, Model Context Protocol (MCP) (Stripe documentation), retrieved 2026-10-28.
- Model Context Protocol, Connect to remote MCP servers, retrieved 2026-10-28.