The problem MCP solves is a multiplication
A model on its own can only write text. To read your calendar, query a database or send a message it needs a tool, and what is an AI agent explains how a model uses tools in a loop. Until late 2024 every app that wanted a tool wrote its own connector for it. Five apps that each wanted twenty tools meant a hundred connectors, each slightly different, each maintained by someone.
MCP is a standard shape for the connector. The tool side is packaged once as a server. The app side implements the client once. After that, any client can use any server: five apps plus twenty tools is twenty-five pieces of software instead of a hundred, and adding a twenty-first tool is one server rather than five connectors. USB is the usual analogy, and it's accurate: the value is the plug, not any one device.
Anthropic published the specification, reference servers and software kits on 25 November 2024 as an open standard, and the reason it spread is the reason any plug spreads. Once a few hundred servers existed for the common things, GitHub, Slack, Postgres, the file system, a browser, an app that spoke MCP got all of them for free, and an app that didn't had to explain why.
A server offers three things, and the model triggers only one of them
An MCP connection has two ends. The host application, a chat app, an editor or an agent runtime, runs a client. The client connects to one or more servers, asks each what it offers, and passes the answers to the model. A server can offer three kinds of thing.
| Primitive | What it is | Who triggers it | Example |
|---|---|---|---|
| Tool | A function the model can call, with typed inputs | The model, when it decides the task needs it | search_issues(repo, query), send_email(to, body) |
| Resource | Data the app can read into context | The app or the user, as context for a turn | A file's contents, a database schema, a web page |
| Prompt | A reusable instruction template | The user, from a menu | "Review this pull request against our style guide" |
The wire format is JSON-RPC 2.0, the same request-and-response convention editors use to talk to language servers. That isn't an accident. MCP borrowed the Language Server Protocol's design, where one standard let every editor support every language. A server can run locally as a subprocess talking over standard input and output, or remotely over HTTP. The March 2025 revision added a streamable HTTP transport and OAuth-based authorisation for remote servers; the June 2025 revision tightened that authorisation and added a way for a server to ask the user a question mid-task, called elicitation.
The part that matters for safety is where the loop lives. The server exposes capabilities; the client and the model decide when to use them. A well-built host asks you before a tool with side effects runs, and what is prompt injection explains why that confirmation is the whole security model once a tool can read untrusted text and act on it.
OpenAI, Google and Microsoft adopted MCP within six months
A standard is only a standard once competitors use it. The timeline below is the adoption record from the announcement to the handover of governance.
OpenAI's adoption in March 2025 was the turning point, because it meant OpenAI and Anthropic had agreed on the plug, and Google followed within two weeks. Microsoft announced MCP support in Windows at Build in May 2025. By December 2025 Anthropic handed the specification to the Agentic AI Foundation, a project under the Linux Foundation co-founded with OpenAI and Block, which is the moment a company's protocol becomes nobody's and therefore everybody's.
What MCP means if you only use a chat app
Directly, little. MCP is plumbing between programs, and it shows up as a feature in the products built on it: a "connectors" menu in a chat app, a tool list in a coding assistant, a way to point your editor at your company's database. It arrives inside an app that speaks it, usually as a connectors menu, and there's nothing for you to install.
Indirectly, it changes what to expect from any AI product. Before MCP, a tool integration was a reason to pick one app over another, because each app built its own. After it, integrations are a commodity and the questions that remain are which model answers, what it costs, and how the app handles the confirmation step before a tool acts. Whizi doesn't expose MCP servers or a developer API. Its tools are the built-in ones, web search and document upload, and its differentiation is the model layer, 280+ models behind one picker at a fixed credit price per message. For agent work that needs your own tools, an MCP-capable agent runtime is the right product and a chat workspace is not.
The one question to carry into any product that advertises MCP: what happens between the model deciding to call a tool and the tool running? If the answer is "nothing", the model can act on whatever text it last read, including text an attacker put in front of it. If the answer is "you approve it", the protocol is doing its job.
- Read "supports MCP" as "can use any MCP server", which is a plug, not a capability in itself
- Ask what confirmation sits between a model choosing a tool and the tool running before connecting anything with side effects
- Prefer remote servers with OAuth for anything shared; local subprocess servers are for your own machine
- Treat the model as the decider and the server as the menu; the server never calls itself
- For chat without tools, MCP changes nothing; pick on the model and the price
Frequently asked questions
What does MCP stand for?
Model Context Protocol. Anthropic published it on 25 November 2024 as an open standard for connecting AI applications to external tools and data. The name describes the job: a protocol for getting context, meaning tools, files and data, to a model in a standard way.
What is an MCP server?
A program that exposes tools, resources or prompts in the MCP format. It might wrap a database, a file system, GitHub, a calendar or a browser. Written once, it works with any MCP client, so a server for your company's data can be used from a chat app, a coding assistant and an agent runtime without three integrations.
Is MCP only for Claude?
No. Anthropic created it, but OpenAI adopted it in March 2025, Google DeepMind committed Gemini to it in April 2025, and Microsoft added it to Windows. In December 2025 the specification moved to the Agentic AI Foundation under the Linux Foundation, co-founded by Anthropic, OpenAI and Block. It's vendor-neutral by design and by governance.
Is MCP the same as function calling or tool use?
Function calling is the model's ability to request a tool. MCP standardises how the application finds, describes and runs those tools. A model with function calling still needs someone to wire up each tool; MCP is the wiring, so the same server plugs into any model that can call functions.
Does Whizi support MCP?
No. Whizi is a chat workspace whose tools are built in, web search and document upload, and it does not expose MCP servers or a developer API. Its value is the model layer, 280+ models behind one picker at a fixed credit price per message. For agent work that needs custom tools, use an MCP-capable agent runtime and bring the model of your choice.
