Model Context Protocol (MCP)

MCP is a standard that allows AI agents to connect to external data and tools. It turns your AI from a text generator into a capable assistant that can do things.

What can MCP do?

Read Docs: Connect to a documentation server to read latest API reference.
Access Database: Query your PostgreSQL or Supabase database directly to understand your schema.
Manage Projects: Create Linear tickets or GitHub issues.
Design: Read Figma files to generate code that matches the design pixel-perfectly.

Setting up MCP in Cursor

You can add MCP servers to Cursor to give your AI new capabilities.

Open Cursor Settings (Cmd + , or Ctrl + ,).
Go to General > MCP.
Click "Add new MCP server".

Example: Adding a Local File Server

This allows the AI to read files outside the current project (e.g., a shared library).

Name: shared-lib
Type: command
Command: npx -y @modelcontextprotocol/server-filesystem /path/to/shared/lib

Example: Adding a Database Inspector

Allows the AI to run read-only queries against your SQLite database.

Name: sqlite-inspector
Type: command
Command: uvx mcp-server-sqlite --db-path ./my-database.db

Example: Adding a Server in Claude Code

Claude Code reads MCP servers from a config file rather than a settings UI. Add them under an mcpServers key, where each entry names the command and arguments used to launch the server:

Json

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/shared/lib"]
    },
    "sqlite-inspector": {
      "command": "uvx",
      "args": ["mcp-server-sqlite", "--db-path", "./my-database.db"]
    }
  }
}

Restart the agent after editing the config so it spawns the new servers on startup.

Verifying a Server Works

Once a server is connected, confirm the agent can actually see its tools before you rely on it:

Ask the agent to list its available tools. The MCP server's tools should appear by name.
Give it a read-only task first — "describe the schema of the users table" — so you can sanity-check the connection without risking a write.
If nothing shows up, check the command runs in your own shell. Most failures are a missing binary (npx, uvx) or a wrong path, not the protocol itself.

Security Notes

MCP servers run with the permissions of whatever command launches them. Two rules keep this safe:

Prefer read-only access. Connect database servers with a read-only role so the agent can inspect schema and data but cannot mutate it.
Scope filesystem servers tightly. Point them at a specific directory, never your home folder or the root of your machine.
Review the source of any third-party MCP server before running it, the same way you would any other dependency.

Integrating MCP servers is the ultimate "Context Engineering". Instead of manually pasting schema definitions into the chat, you give the AI direct access to the database to "see" for itself. The payoff is concrete: the model stops guessing at your column names and starts generating code against the schema you actually have — which is the single biggest source of avoidable AI hallucinations in a real project.

MCP Servers