Skip to main content
VibeStacks

Loading the search index…

Entry

MCP

The Model Context Protocol connects external capabilities such as browsers, databases, and docs search to any AI tool that supports it, through one interface.

Last reviewed
Website
modelcontextprotocol.io

Good for

  • Letting the agent do what it otherwise cannot: drive a browser, query a database, read designs or tickets
  • External capabilities you want to set up once and share across tools
  • Cases that need live data (current docs, production status) rather than model memory

Not for

  • Telling the agent your project conventions (use AGENTS.md)
  • A fixed sequence of steps (use Skills, which are lighter)
  • Third-party servers of unknown origin: the model reads their tool descriptions and results

MCP (Model Context Protocol) is an open protocol based on JSON-RPC. Anthropic released it in November 2024 and donated it to the Agentic AI Foundation under the Linux Foundation in December 2025. It defines how an AI application (the host) connects to external MCP servers.

A server can offer three things

  • Tools: operations the model can call, such as "open a page and take a screenshot" or "run a query".
  • Resources: data that can be read, such as files, schemas, or recent tickets.
  • Prompts: parameterized templates the user can invoke like commands.

Write an MCP server once, and Cursor, Claude Code, and any other MCP-capable tool can use it.

When it is worth connecting

Ask yourself: is the agent failing because it does not know something, or because it cannot do something? If it does not know your conventions, write AGENTS.md. Only when it cannot act (for example, it cannot open a browser to check a page) should you add MCP. For frontend work the most useful one is browser automation: the agent opens the page, clicks, takes screenshots, and verifies that a change really works.

Security

An MCP server's tool descriptions and results go into the model's context, and the model may follow instructions hidden in them. Install only servers you trust, and keep human confirmation before tools that write (sending messages, changing data).

Configuration

Most tools declare servers in a JSON file with the same basic shape: each entry under mcpServers gives a command and its arguments. Cursor reads .cursor/mcp.json in the project; Claude Code reads .mcp.json in the project root. The souvenir below configures Playwright's browser automation server; put it where your tool expects it.

Souvenir

Take it home to your AI

Drop this mcp.json into your project, or paste it straight to your AI coding assistant.

mcp.json
{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    }
  }
}

Related

  • Article

    The pieces of context you give AI

    AGENTS.md sets project rules, Rules handle local scope, MCP adds external capabilities, and Skills hold know-how loaded on demand.

  • Entry

    Skills

    Package how to do a kind of task as a folder. The agent remembers only its name and description until a task calls for the full text.

  • EntryTop pick

    Cursor

    An AI IDE built on VS Code. Completion, chat, the agent, and your code live in one window, and every change arrives as a diff.