Guide

MCP for Beginners

A practical introduction to connecting tools to AI agents.

Author: Manus AI - January 12, 2026

Introduction

The Model Context Protocol (MCP) is an open standard for connecting AI clients to external tools and data sources. It gives models a consistent way to use tools (actions), access resources (data), and run prompts (templates) without custom per-app integrations.

With MCP, a single server can work with multiple clients that support the protocol. The result is reusable integrations, clearer permissions, and workflows that are easier to standardize across teams.

TL;DR

MCP is an open protocol for connecting AI clients to tools and dataServers expose tools, resources, and promptsMessages follow JSON-RPC 2.0Start with a small, read-only serverUse least-privilege credentials and separate environmentsPrefer authenticated transports for remote servers

What is MCP?

A standard protocol for tool connections

MCP defines a consistent, JSON-RPC-based interface that clients and servers can use to exchange tools, resources, and prompts. That standardization is what makes integrations reusable.

How it differs from one-off plugins

Rather than building bespoke integrations per client, MCP lets a server expose capabilities once and makes them available to any MCP-capable client. This reduces maintenance and improves consistency across tools.

When MCP is the right fit

Use MCP when your workflows need data or actions from outside your AI client: files, tickets, databases, email, calendars, or web sources. If copy-paste is part of the workflow, MCP is usually the next step.

Core concepts

Servers

Servers expose capabilities to clients. They can provide tools (actions), resources (data), and prompts (templates) over the MCP protocol.

Tools

Model-controlled actions like API calls, file writes, or database queries. Clients can apply permission gates before execution.

Resources

Client-attached context such as file contents, database records, or API responses. Resources are typically read-only.

Prompts

User-controlled templates that standardize common workflows. In Claude Code, MCP prompts can appear as slash commands.

Getting started

Confirm MCP support in your client

Use a client that supports MCP servers, prompts, and tools. Claude Code includes MCP support and can discover prompts dynamically.

Choose a first server with low risk

Start with a local, read-only server like filesystem or a sandbox database. Avoid write permissions until the workflow is stable.

Connect the server

Add the server using your client’s configuration or CLI. In Claude Code, the MCP CLI supports local, project, and user scopes.

Test the connection

List available tools and prompts, then run a simple read-only action to confirm output shape and permissions.

Lock down permissions

Use least privilege, audit logs, and separate credentials for dev vs production. Add human review for write actions.

Use cases

Development

Codebase analysis

Connect filesystem and version control servers to map architecture, find ownership, and review changes quickly.

filesystemgit
Development

Database exploration

Query schemas and run read-only analyses directly from your AI client without context switching.

postgresmysqlsqlite
Productivity

Email and calendar

Summarize inboxes, prepare meeting briefs, and draft replies with explicit approval steps.

emailcalendar
Research

Web search and retrieval

Gather sources, extract key facts, and build structured research briefs without copy-paste.

web searchfetch
Business

Reporting and analytics

Combine CRM, spreadsheets, and databases to generate weekly reporting and KPI rollups.

sheetscrmdatabase
Operations

Workflow automation

Chain tools like tickets + chat + docs to automate triage, handoffs, and status updates.

ticketingchatdocs

Common issues

Server not appearing

  • Verify the server command and working directory
  • Restart the client to reload MCP configuration
  • Check server logs for startup errors

Permission errors

  • Confirm credentials and scopes
  • Use read-only access while testing
  • Verify the client’s allowed directories

Slow or failing requests

  • Reduce query size or add pagination
  • Check upstream API limits and quotas
  • Add caching or batching in the server

Best practices

Treat MCP servers like production integrations: secure, versioned, and monitored
Start with read-only tools and add write access only with review steps
Document prompt templates and expected outputs for repeatability
Separate dev and prod credentials to reduce risk
Prefer scoped access and data filters over broad permissions

Common server categories

File & Data

FilesystemRead project files and docs
GitInspect history, diffs, and blame
PostgreSQLQuery relational data

Web & APIs

HTTP FetchRetrieve web pages and APIs
SearchFind sources and references
Browser automationScripted browsing or scraping

Productivity

EmailSummaries, drafts, triage
CalendarAvailability and scheduling
Docs/WikiRead and update knowledge

Operations

TicketsTriage and update issues
ChatSummaries and announcements
CRMAccount context and reporting