What is MCP? The Model Context Protocol, explained for non-engineers

MCP, the Model Context Protocol, is an open standard for connecting AI agents to tools and data. A tool's maker publishes an MCP server that describes what an agent can do there (search email, create an issue, update a deal) and any AI app that speaks MCP can use it. Think of it as USB-C for AI tools: one plug that fits everywhere.

Anthropic introduced MCP in November 2024. Since then OpenAI, Google, Microsoft, and most AI app makers have adopted it, and thousands of tools ship MCP servers.

Why MCP exists

Before MCP, every AI app had to build its own integration with every tool. Ten apps and a hundred tools meant a thousand custom integrations, most of them half-finished.

MCP flips that. A tool builds one MCP server; every MCP-speaking app can use it. An app supports MCP once; it can use every MCP server. That's why the number of tools available to agents grew so fast.

How MCP works

An MCP server offers three kinds of things to an AI app:

• Tools: actions the agent can take, like send_email or create_calendar_event. Each comes with a description and the inputs it needs, so the model knows when and how to call it.
• Resources: data the agent can read, like files or records.
• Prompts: ready-made instructions a server suggests for common tasks.

When an agent runs, the app tells the model which tools exist. The model decides to call one, the app sends the call to the MCP server, and the result comes back to the model. That loop is what makes an AI agent an agent.

Local vs remote MCP servers

There are two ways to run an MCP server:

• Local servers run on your own computer, started by the app. Developers use these a lot, for example to give a coding assistant access to files.
• Remote servers run on the internet at an address like https://mcp.example.com. You connect over HTTPS and usually sign in with OAuth, the same "Sign in with Google" style flow you already know.

Phone apps like Stormi use remote servers, because a phone can't run servers for you and because remote servers keep working when your phone is off.

How signing in works

Most remote MCP servers need you to sign in, so the agent acts as you. The MCP spec uses OAuth for this:

1. The app sends you to the tool's own sign-in page.
2. You sign in there and approve access.
3. The tool gives the app a token. Your password never leaves the tool's page.

In Stormi, that token is encrypted on our server and never sent to your iPhone, and bots never see your password.

Annotations: how an agent knows what's risky

MCP lets a server mark a tool as read-only or destructive. Stormi uses those annotations to sort every tool into read, write, or irreversible, and irreversible calls (send, pay, delete, submit) wait for your approval. When a server doesn't annotate its tools, the verb in the tool's name decides: delete_ and send_ are treated as irreversible.

Using MCP without writing code

You don't need to know any of this to benefit from it. In Stormi:

1. Open a bot and tap Plugins.
2. Turn on a plugin from the catalog, like GitHub or HubSpot, and sign in.
3. Or tap Add Plugin and paste the address of any remote MCP server. Or just ask a bot: "set up the Acme MCP server." It reads Acme's docs, finds the address, and adds it once you approve.

See the full list on the integrations page.

Is MCP safe?

MCP itself is plumbing. Safety depends on the app and the server:

• Only connect servers from tools you already trust. A server's tool descriptions and results reach the model.
• Prefer apps that ask before irreversible actions.
• Prefer apps that treat tool results as data, not instructions, so a malicious email can't steer the agent. More in prompt injection, explained.