Host
The AI app you use. It runs the model, shows you results and decides which tool calls need your approval.
Free explainer
MCP (Model Context Protocol) is an open standard that lets AI apps such as Claude, ChatGPT, Cursor and VS Code connect to outside tools and data. Think of it as a USB-C port for AI: any app that speaks MCP can plug into any MCP server, whether that server reads your files, queries a database or posts to Slack.
Below, step through what actually happens when an AI uses MCP, check whether your own task needs it, and browse 31 real MCP servers. No sign-up; everything runs in your browser.
Pick a task, then step through what happens between you, the AI app and the MCP server.
Step 1 of 7
You type: "Summarize notes.md in my Projects folder." The AI app (the host) cannot do this alone. It needs a tool.
Transport in this example: stdio (the server runs as a local program on your computer).
The AI app you use. It runs the model, shows you results and decides which tool calls need your approval.
A connector inside the host. Each client keeps one connection to one server and passes requests back and forth as JSON-RPC messages.
A program that offers tools (actions), resources (data to read) and prompts (templates). It runs locally over stdio or remotely over Streamable HTTP.
Choose a common task or answer the four questions yourself.
Verdict
The AI needs to choose when to read or change data in a cloud service. Many vendors host an official remote MCP server; you add its URL and sign in with OAuth, and every MCP-capable AI app can use the same server.
How to set it up: Add the server URL as a custom connector or in your MCP config, then sign in when the app opens the OAuth window.
Example: HubSpot MCP server
Official or reference servers you can add today. Local servers run on your computer; remote servers are hosted by the vendor and you sign in with OAuth.
31 servers
Read, write and search files in folders you allow
Reference
Read history, diffs and branches of a local repository
Reference
Fetch a web page and convert it to Markdown
Reference
Keep a knowledge graph the AI can recall later
Reference
Current time and time zone conversion
Reference
Step-by-step reasoning scratchpad
Reference
Issues, pull requests, code search and Actions
Developer
Drive a real browser: click, type, read pages
Developer
Inspect, debug and profile pages in Chrome
Developer
Look up errors, stack traces and releases
Developer
Current library docs pulled into the prompt
Developer
Create, find and update issues and projects
Work & docs
Search, read and edit pages and databases
Work & docs
Jira issues and Confluence pages
Work & docs
Search channels, read threads, send messages
Work & docs
Tasks, projects and goals
Work & docs
Search and read files stored in Box
Work & docs
Query tables, manage schema and projects
Data
Serverless Postgres: run SQL, branch databases
Data
Query collections and manage Atlas clusters
Data
Google's server for Postgres, MySQL, BigQuery and more
Data
Customers, payments, refunds and docs search
Business apps
Invoices, orders and transactions
Business apps
Payments, catalog, orders and customers
Business apps
Contacts, companies and deals in your CRM
Business apps
Actions in thousands of connected apps
Business apps
Web, news and image search
Web & research
AI-oriented web search and page contents
Web & research
Search official Microsoft and Azure docs
Web & research
Workers, DNS, logs and docs for your account
Cloud & design
Design context and code from Figma frames
Cloud & design
| MCP | Direct API | Function calling | Plugins | |
|---|---|---|---|---|
| What it is | An open protocol that connects AI apps to tools and data | A service's own interface for programs | A model feature: the model returns a structured tool call | A product-specific add-on system (for example ChatGPT plugins, 2023 to 2024) |
| Who calls it | The AI app, when the model chooses a tool | Your code, on a fixed path | Your code, after the model asks for a call | One AI product |
| Works across AI apps | Yes: build one server, use it in Claude, ChatGPT, Cursor, VS Code | Each app needs its own integration code | Tool definitions are written per app | No, tied to one product |
| Tool discovery | Built in: tools/list at run time | Read the docs | You pass the tool list in each request | Manifest file |
| Auth | OAuth for remote servers; local servers use your machine | API keys or OAuth | Whatever your code uses | Set by the product |
| Best for | Letting assistants and agents use many tools safely | Scripts, back ends and fixed automations | Tools inside one app you are building | Largely replaced by MCP and app connectors |
They work together: an AI app often uses the model's function calling to decide on a tool, MCP to reach it, and the server uses the service's API to do the work.
A local server runs as a program with your user's rights. Prefer the vendor's official server or the reference servers, and check who publishes a server before you add it.
Limit folders, use read-only database logins and narrow OAuth scopes. Turn off tools you do not need.
Let the AI app ask before any tool that sends, deletes, pays or posts. Read the arguments before you click Allow.
Text a tool returns, such as a web page, email or issue, can contain instructions aimed at the model. Do not let untrusted content trigger tools with write access.
Put API keys in environment variables or your app's secret store, not in prompts or committed config files.
List your connected servers now and then, revoke OAuth grants you no longer use, and update servers like any other software.
Choose an official server for the app you want to connect, such as GitHub, Notion or Slack, from the vendor docs or the MCP Registry.
For a remote server, paste its URL as a custom connector. For a local server, add its command to the app's MCP config file.
Complete the OAuth sign-in or set the API key, and grant only the scopes, folders or tools you need.
Ask the AI to use it in plain English. Approve tool calls that change data after checking the arguments.
MCP (Model Context Protocol) is an open standard for connecting AI applications to external tools and data. An AI app such as Claude, ChatGPT, Cursor or VS Code connects to MCP servers, and each server offers tools the model can call, such as reading a file, querying a database or posting to Slack.
Anthropic introduced MCP as an open-source protocol on November 25, 2024. On December 9, 2025, Anthropic donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI.
No. MCP is supported by many AI apps, including ChatGPT, Visual Studio Code, Cursor and Claude, and by agent frameworks. A server built once works in any MCP-capable app.
An MCP server is a program that exposes tools, resources and prompts to AI apps using MCP. Local servers run on your computer and talk over stdio; remote servers run on the web and use Streamable HTTP, usually with OAuth sign-in.
MCP itself is a protocol; safety depends on which servers you connect and what they can do. Use servers from trusted publishers, grant the least access, keep approval on for actions that change things, and be careful with untrusted content that could carry prompt injection.
No. In most AI apps you add an MCP server by pasting its URL as a connector or by adding a short config entry, then signing in. You only need to code to build your own server.
An API is how programs talk to one service. MCP is a standard layer on top that lets an AI app discover a server's tools at run time and call them when the model decides to. Many MCP servers simply wrap a service's API.
Checked 2026-11-14 against the official MCP docs (current spec 2026-07-28) and Anthropic's launch post and donation announcement.
Tornic turns the Claude Code, Codex or Cursor CLI you already pay for into deterministic multi-step workflows, so your agents can use the MCP servers you connect the same way every run.
Explore Tornic