MCP Guide 2026: Connect Claude to Any Tool
Model Context Protocol (MCP) lets Claude connect to databases, APIs, and tools in 2026. Complete setup guide with 10 server examples and code.
Model Context Protocol (MCP) is the open standard that lets Claude connect to any external tool — databases, APIs, file systems, code repositories — without custom integration work. Released by Anthropic in November 2024 and now adopted across the AI ecosystem, MCP turned Claude from a standalone chat interface into a programmable agent that operates inside your actual infrastructure. This guide covers how MCP works, how to set it up in Claude Code in under ten minutes, and which of the 200+ available MCP servers are worth installing first.
TL;DR
- MCP is an open protocol that gives Claude structured access to external tools and data sources.
- Claude Code supports MCP natively — add servers via
claude mcp addin one command. - The host (Claude Code) connects to MCP servers (your tools) through a client layer.
- Over 200 MCP servers are publicly available covering databases, APIs, browsers, and dev tools.
- No API key sharing or custom plugin code required — servers run locally or on your infrastructure.
What Is MCP and Why It Matters
Before MCP, connecting an AI assistant to a database or API required writing custom tool definitions for every integration. Each model had its own format. Every new tool meant new glue code. MCP solves this with a single, standardized protocol that any model and any tool can speak.
Think of MCP like USB-C for AI tools. Instead of a different cable for every device, you have one standard port. Any MCP-compatible server (the “device”) plugs into any MCP-compatible client (the “port”) without custom adapters.
The practical result: Claude can query your Postgres database, read your GitHub issues, search your Notion docs, and run browser automation — all in a single conversation, with no code you wrote to connect them.
MCP Architecture
MCP uses a three-component model:
| Component | Role | Example |
|---|---|---|
| Host | The AI application that manages connections | Claude Code, Claude Desktop |
| Client | Protocol handler inside the host (1:1 per server) | Built into Claude Code |
| Server | External process exposing tools and resources | @modelcontextprotocol/server-postgres |
Each MCP server exposes three primitives:
- Tools — Functions Claude can call (e.g.,
query_database,create_issue) - Resources — Data Claude can read (e.g., file contents, API responses)
- Prompts — Pre-built prompt templates the server provides
Communication uses JSON-RPC 2.0 over stdio (local) or HTTP with Server-Sent Events (remote). Local servers are the most common and run entirely on your machine — no data leaves your environment.
Setup: Adding MCP Servers to Claude Code
Prerequisites
- Claude Code installed (
npm install -g @anthropic-ai/claude-code) - Node.js 18+ or Python 3.10+ depending on the server
- A Claude Pro or Max subscription
Method 1: CLI (Recommended)
# Add a server by package name claude mcp add @modelcontextprotocol/server-filesystem # Add with custom arguments claude mcp add @modelcontextprotocol/server-postgres \ --env DATABASE_URL="postgresql://localhost/mydb" # List installed servers claude mcp list # Remove a server claude mcp remove server-filesystem Method 2: Config File
MCP servers are stored in ~/.claude/mcp.json. You can edit this directly:
{ "mcpServers": { "filesystem": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/projects"], "env": {} }, "postgres": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-postgres"], "env": { "DATABASE_URL": "postgresql://localhost/mydb" } } } } Verify the Connection
# Start Claude Code and check tool availability claude > /mcp # Claude will list all connected servers and their available tools 10 MCP Servers Worth Installing in 2026
{ "data_updated": "2026-04-10", "total_available_servers": 200, "servers": [ { "name": "Filesystem", "package": "@modelcontextprotocol/server-filesystem", "category": "File System", "tools": ["read_file", "write_file", "list_directory", "search_files"], "use_case": "Read and write files in specified directories", "install": "claude mcp add @modelcontextprotocol/server-filesystem /path/to/dir", "local_only": true, "official": true }, { "name": "PostgreSQL", "package": "@modelcontextprotocol/server-postgres", "category": "Database", "tools": ["query", "list_tables", "describe_table"], "use_case": "Read-only SQL queries against your Postgres database", "install": "claude mcp add @modelcontextprotocol/server-postgres --env DATABASE_URL=...", "local_only": true, "official": true }, { "name": "GitHub", "package": "@modelcontextprotocol/server-github", "category": "Code Repository", "tools": ["list_repos", "get_issue", "create_issue", "search_code", "get_pr"], "use_case": "Read/write GitHub issues, PRs, and code search", "install": "claude mcp add @modelcontextprotocol/server-github --env GITHUB_TOKEN=...", "local_only": false, "official": true }, { "name": "Brave Search", "package": "@modelcontextprotocol/server-brave-search", "category": "Web Search", "tools": ["web_search", "local_search"], "use_case": "Real-time web search without leaving Claude Code", "install": "claude mcp add @modelcontextprotocol/server-brave-search --env BRAVE_API_KEY=...", "local_only": false, "official": true }, { "name": "Puppeteer", "package": "@modelcontextprotocol/server-puppeteer", "category": "Browser Automation", "tools": ["navigate", "screenshot", "click", "fill_form", "extract_content"], "use_case": "Headless browser control for scraping and E2E testing", "install": "claude mcp add @modelcontextprotocol/server-puppeteer", "local_only": true, "official": true }, { "name": "SQLite", "package": "@modelcontextprotocol/server-sqlite", "category": "Database", "tools": ["query", "list_tables", "create_table", "insert"], "use_case": "Full read/write access to local SQLite databases", "install": "claude mcp add @modelcontextprotocol/server-sqlite /path/to/db.sqlite", "local_only": true, "official": true }, { "name": "Slack", "package": "@modelcontextprotocol/server-slack", "category": "Communication", "tools": ["list_channels", "post_message", "get_thread", "search_messages"], "use_case": "Read Slack channels and post messages from Claude", "install": "claude mcp add @modelcontextprotocol/server-slack --env SLACK_TOKEN=...", "local_only": false, "official": true }, { "name": "Google Maps", "package": "@modelcontextprotocol/server-google-maps", "category": "Location", "tools": ["geocode", "reverse_geocode", "search_places", "get_directions"], "use_case": "Location data and directions for apps with geo features", "install": "claude mcp add @modelcontextprotocol/server-google-maps --env MAPS_API_KEY=...", "local_only": false, "official": true }, { "name": "Sentry", "package": "@modelcontextprotocol/server-sentry", "category": "Observability", "tools": ["get_issues", "get_event", "list_projects", "resolve_issue"], "use_case": "Query Sentry errors directly from your coding session", "install": "claude mcp add @modelcontextprotocol/server-sentry --env SENTRY_TOKEN=...", "local_only": false, "official": true }, { "name": "Memory", "package": "@modelcontextprotocol/server-memory", "category": "Persistence", "tools": ["store", "retrieve", "list", "delete"], "use_case": "Persistent key-value memory across Claude sessions", "install": "claude mcp add @modelcontextprotocol/server-memory", "local_only": true, "official": true } ] } Real Workflow Examples
Example 1: Debug a Production Error End-to-End
With Sentry + GitHub + Filesystem MCP installed:
You: "Find the most recent unresolved Sentry error in the payments service, locate the relevant code, and suggest a fix." Claude: 1. [Sentry] Queries latest unresolved issues → finds NullPointerException in checkout.py:142 2. [GitHub] Searches codebase for checkout.py → retrieves current implementation 3. [Filesystem] Reads local version with recent uncommitted changes 4. Analyzes root cause and writes a patch Example 2: Generate a Report from Live Data
With PostgreSQL + Filesystem MCP:
You: "Query the orders table for last week's revenue by product category and save it as a CSV to /reports/weekly-revenue.csv" Claude: 1. [PostgreSQL] Runs aggregation query 2. [Filesystem] Writes CSV to specified path 3. Reports row count and total revenue figure Example 3: Automate a Web Form
With Puppeteer MCP:
You: "Go to our staging environment at localhost:3000, log in with test@example.com/password123, and verify the checkout flow works." Claude: 1. [Puppeteer] Navigates to localhost:3000 2. Fills login form and submits 3. Walks through checkout steps 4. Reports pass/fail with screenshot Security Considerations
MCP servers run with the permissions of the process that starts them. Before installing any server:
- Review the package source — prefer the official
@modelcontextprotocol/namespace - Scope filesystem access — pass only the directories Claude needs, not
~/ - Use read-only database connections — create a dedicated read-only DB user for MCP
- Rotate tokens regularly — MCP server tokens stored in
mcp.jsonshould be short-lived where possible - Audit tool calls — Claude Code logs all MCP tool invocations; review them in
.claude/logs/
Frequently Asked Questions
Does MCP work with Claude.ai (the web interface)?
MCP is currently supported in Claude Code (CLI) and Claude Desktop. The claude.ai web interface does not support custom MCP servers as of April 2026. For MCP-enabled workflows, Claude Code is the primary interface.
Can I build my own MCP server?
Yes. Anthropic provides SDKs in TypeScript and Python. A minimal server takes about 50 lines of code. The TypeScript SDK (@modelcontextprotocol/sdk) is the more mature option. See the official spec at modelcontextprotocol.io for the full schema.
Are MCP servers safe to use with proprietary code?
Servers marked local_only: true (like Filesystem, SQLite, Puppeteer) run entirely on your machine — no data is sent to any external service beyond what Claude Code itself sends to Anthropic’s API. For servers that connect to external APIs (GitHub, Slack), data flows through those services’ standard APIs under their existing privacy policies.
How many MCP servers can I run simultaneously?
There is no hard limit. In practice, 5–10 servers covers most workflows. Each server spawns a separate process; memory usage scales linearly with the number of active servers. Unused servers can be disabled in mcp.json without uninstalling.
Related Posts
- Best AI Coding Tools 2026: Claude Code vs Cursor vs Copilot — full tool comparison with benchmark data
- Claude Code Getting Started 2026: Install to First Agent — step-by-step setup guide for Claude Code
- Best LLMs for Coding 2026: Benchmark Results — model comparison for developer use cases
Last updated: 2026-04-10