How the Model Context Protocol (MCP) works: a plain-English guide for support and ops teams

The problem before MCP

Say you want your AI support agent to look up an order in Shopify, check a subscription in Stripe, and post to a Slack channel when it gets stuck. Before MCP, that meant three different vendor plugin systems, three different auth flows, three different upgrade cycles, and three different places to find out that a change broke your bot.

And that's if the plugins existed. If not, you paid an engineer to write a bespoke integration in your AI vendor's proprietary format, which meant re-doing the work if you ever switched vendors.

The MCP idea

MCP is a small JSON-RPC protocol that says: any program that exposes tools can implement one server. Any AI application that wants to use tools implements one client. The two negotiate, the client asks “what tools do you have?”, the server answers with a list of names, descriptions, and input schemas, and the AI decides which to call.

Because it's a standard rather than a marketplace, the integrations you write once work everywhere the standard is spoken. Anthropic's Claude Desktop and Cursor and OpenAgent and (soon) most enterprise AI tools all speak it. So do the community servers that already exist for Shopify, Stripe, HubSpot, Notion, Postgres, GitHub, and dozens more.

The moving parts

• Server. A program that exposes tools (also resources and prompts, but tools are the main event). Runs somewhere reachable by HTTPS, or as a local subprocess for desktop AI clients.
• Tool. A named function the server offers, with a description in natural language and a JSON schema for its arguments. The LLM reads the description to decide when to call it.
• Client (or “host”). An AI application that connects to servers, gets their tool catalogs, and invokes tools during a conversation. OpenAgent is a client.
• Transport. How the client and server talk. The current standard is Streamable HTTP; the older HTTP-plus-SSE variant is deprecated. Desktop clients also use stdio (subprocess) which OpenAgent deliberately does not support, for the multi-tenant safety reasons in the MCP feature page.
• Auth. Usually a bearer token (a static API key the client sends on every request). OAuth 2.1 is the spec direction for 2026 and after.

What changes for you as a buyer

Three practical shifts:

1. You start asking “does this AI vendor speak MCP?” If yes, you inherit the growing catalog of MCP servers. If no, you're back to bespoke integrations forever.
2. Integration lock-in goes down. Switching AI platforms doesn't mean re-writing your Shopify or Stripe integration. If both platforms speak MCP, the same server works with both.
3. Custom actions become 100 lines of code, not a project. If your team wants the AI to hit an in-house API (“check whether this customer is a VIP in our loyalty DB”), your engineers write a small MCP server, deploy it behind HTTPS, and the tenant admin pastes the URL into OpenAgent. No plugin submission, no vendor feature request.

What MCP doesn't change

MCP is a plumbing standard. It doesn't make the AI smarter, doesn't reduce hallucinations, and doesn't magically know when to call which tool. The LLM still reads the description you (or the server author) wrote and makes a judgment call. If the description is vague, the AI will under-use or misuse the tool. Reviewing and rewriting descriptions with descriptionOverride is one of the higher-leverage things an admin can do after wiring up a new server.

MCP also doesn't solve the trust problem. A malicious or buggy server can lie in its tool descriptions and the AI has no way to know. That's why the client's safety rails (approval, drift detection, audit) matter as much as speaking the protocol at all.

Where to go from here

For the buyer-side pitch on why we shipped MCP the way we did, read the MCP feature page. For the canonical protocol spec and the community server directory, Anthropic maintains modelcontextprotocol.io. For the workspace-admin walkthrough of registering a server and approving tools in OpenAgent, the in-product help centre has it once your tenant is enabled.