MCP Servers for AI Coding: The Developer's Guide

MCP stands for Model Context Protocol, an open standard created by Anthropic that defines how AI models connect to external tools and data. Think of it as USB for AI agents: any MCP-compatible agent can use any MCP server, no bespoke integration required.

An MCP server exposes tools (functions the agent can call) and resources (data it can read). The agent doesn't care whether it's talking to a database, a web API, or the filesystem, it calls tools through the same interface. Claude Code, Codex, and most modern agentic IDEs speak MCP, and hundreds of servers exist as of 2026.

Without MCP, an agent's knowledge of the outside world is frozen at its training cutoff. It can read your files and run shell commands, but it can't check whether a library shipped a breaking change last week, or verify what a GitHub issue actually says. With the right servers, that changes:

These are the servers with the highest day-to-day value for developers running coding agents, whether through Claude Code directly or an agent harness that manages them for you. Each install command runs the server locally over stdio.

npx @modelcontextprotocol/server-brave-search

npx @modelcontextprotocol/server-fetch

npx @modelcontextprotocol/server-postgres

npx @modelcontextprotocol/server-sqlite

npx @modelcontextprotocol/server-github

npx @modelcontextprotocol/server-filesystem

npx @modelcontextprotocol/server-memory

npx @modelcontextprotocol/server-playwright

Claude Code manages MCP servers through the claude mcp add command, not by hand-editing a config file. Adding the GitHub server, for example:

# add a server (local scope, stored in ~/.claude.json)
claude mcp add github \
  -e GITHUB_PERSONAL_ACCESS_TOKEN=ghp_your_token_here \
  -- npx -y @modelcontextprotocol/server-github

# add a server for the whole team (project scope, .mcp.json in the repo)
claude mcp add --scope project fetch -- npx -y @modelcontextprotocol/server-fetch

# list what's configured
claude mcp list

Scope decides where the entry lives. Local and user scope are stored in ~/.claude.json, private to your machine. Project scope writes a .mcp.json file at the repo root that you check into git, everyone who clones the project gets the same servers. A project-scoped config looks like this:

// .mcp.json, project scope, checked into the repo
{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_your_token_here"
      }
    },
    "fetch": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-fetch"]
    }
  }
}

Each entry gives the command that starts the server process, its arguments, and any credentials as environment variables. Restart your session after adding a server and its tools become available to the agent.

To be clear: you don't need any of this to ship code with AI agents. Claude Code and Codex are useful with zero MCP servers configured, and so is AIDEN. MCP is an extension layer you add when an agent keeps hitting the same wall, stale docs, unknown schema, invisible CI status.

That inheritance is one reason to prefer a tool that builds on Claude Code over one that replaces it: your MCP investment carries over unchanged. Because MCP is a standard rather than a per-product plugin system, a server you configure today works with whatever MCP-compatible agent you run next year.