MCP (Model Context Protocol) Basics

Standard way to plug tools, data sources, and IDEs into LLM clients — one protocol, many servers.

Interview tip MCP = USB-C for AI tools. Host (Cursor, Claude) connects to MCP servers (GitHub, DB, filesystem).

① What you must know (30 sec)

MCP (Model Context Protocol) is an open standard connecting hosts (Cursor, Claude Desktop, custom apps) to MCP servers that expose tools (actions) and resources (readable data). One server implementation works across any MCP-compatible host — like USB-C for AI integrations.

Analogy: MCP is USB-C for AI tools: one port on the host (IDE), many peripherals (servers) that plug in without custom drivers per app.

② How it works

MCP architecture

MCP Host→↔ protocol→MCP Server→Tools + Resources

Host — LLM client (Cursor, Claude Desktop) that runs the agent loop
Server — process exposing tools (run query) and resources (read file, fetch doc)
Discovery — host lists available tools/resources at connection time
Invocation — model selects tool; host routes call to correct server
Transports — stdio (local) or SSE/HTTP (remote servers)

Community servers exist for GitHub, Postgres, Slack, Brave Search — check official MCP registry before building custom.

③ Step-by-step (hands-on)

Step 1 — Identify host and needs

Cursor or Claude Desktop — list data sources and actions the agent needs (repo, DB, tickets).

Step 2 — Install community servers

Configure mcp.json with server command, args, and env vars. Restart host to discover tools.

Step 3 — Authenticate safely

Pass API tokens via env — not committed. Use read-only DB credentials where possible.

Step 4 — Test tool discovery

Ask agent "what tools do you have?" Verify expected tools appear before real tasks.

Step 5 — Build custom server (optional)

Use MCP SDK (TypeScript/Python) to wrap internal API as tools with JSON schemas.

Step 6 — Harden for team rollout

Allowlist approved servers; document each tool capability; audit logs on sensitive tools.

④ Code / config patterns

Concept	Description	Example
Host	Runs LLM + connects servers	Cursor IDE
Server	Exposes capabilities	github-mcp-server
Tool	Action with side effects	create_issue, run_query
Resource	Readable data URI	file://, db schema
Transport	Connection channel	stdio, SSE

// ~/.cursor/mcp.json (conceptual)
{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "${GITHUB_TOKEN}" }
    },
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres"],
      "env": { "DATABASE_URL": "${DATABASE_URL}" }
    }
  }
}

⑤ Production & pitfalls

Pitfall	Why it hurts	Fix
Over-permissioned servers	Agent deletes prod data	Read-only creds; separate write tools with approval
Unvetted community servers	Malicious code in server process	Allowlist; review server source
Secrets in mcp.json committed	Token leak in git	Env var references only; .gitignore config
Too many tools enabled	Model picks wrong tool	Enable minimal set per project
No transport security (remote)	MITM on SSE	TLS, auth tokens, private network
Confusing MCP with function calling	MCP is cross-app standard; function calling is per-API	Use MCP for reusable integrations

Production tips:

Central IT-approved server catalog for enterprise
Per-team mcp.json templates with least privilege
Log every tool invocation with user and args (redacted)
Sandbox custom servers before org-wide deploy

⑥ Interview / on-the-job Q&A

Question	Answer
What is MCP?	Open protocol for LLM hosts to connect to external tools and data via MCP servers.
Host vs server?	Host runs the AI client; server exposes tools/resources the host can call.
Tool vs resource?	Tool performs an action; resource is addressable read-only data.
Why not custom plugins?	Write once, use in Cursor, Claude, and other MCP hosts.
stdio vs SSE?	stdio for local subprocess servers; SSE/HTTP for remote networked servers.
Security concern?	Servers run with your credentials — treat like installing software with API access.

⑦ Tools & ecosystem

Hosts: Cursor, Claude Desktop, custom MCP clients
SDK: @modelcontextprotocol/sdk (TS), mcp Python
Servers: GitHub, Postgres, filesystem, Brave Search, Slack
Registry: MCP server directory (official/community)

mcptoolscursorintegrationprotocol

⑧ Revision checklist

MCP host configured (Cursor mcp.json or Claude config)
Only required servers enabled
Credentials via environment variables
Read-only access for exploratory DB tools
Tool list verified after host restart
Custom servers code-reviewed before use
Team allowlist of approved MCP servers
Sensitive tool calls audit-logged
Understand difference from in-app function calling