MCP

Before an assistant changes an authorization helper, I want to know which routes depend on it, where its callers live, which revision we're reading, and what the edit assumes about the database. Then I'll look at the patch.

A server earns its place by returning evidence I can inspect. Show me the path to the implementation, let me read the commit that explains a condition, keep the revision attached, and I'll compare that evidence with the proposed change.

With MCP tools, a server exposes operations a model can invoke, the protocol leaves the interaction model open, its guidance calls for a person to be able to deny tool calls, and I would keep that approval step when adding servers.

Choose a server by the missing evidence

These are examples for a change review, with a different job for each category, no ranking between them, and a specific limit to check before trusting the result.

A failed tool call and a refactor aimed at the wrong function need different fixes. Before adding another MCP server, work out what you need to see, what you need to build, and which result you can check yourself.

I would start with the official SDK for server authoring. Use MCP Inspector to examine a call. Add a terminal client when you need a repeatable reproduction, then evaluate repository context against code you know, with a named revision, before trusting it on a change.

Choose by the job in front of you

I build Pharaoh. Its job here is repository context, and I’ll show an actual result below, but first you need a server you can call, a way to inspect the exchange, and a clear idea of which

A shared .mcp.json doesn't make an MCP server read-only. Before adding one, I want to know what it answers, how it authenticates, which projects can see it, and what Claude will load into context.

My shortlist is GitHub for issues and pull requests, Context7 for library documentation, Playwright for browser interaction, and Pharaoh (ours) for repository structure. Install the one that fits your next task. You don't need the whole list.

This guide sticks to Claude Code configuration. For a broader comparison, see MCP servers for every client.

Choose a scope before adding the server

Claude Code separates where a configuration applies from what the server can do. Its MCP documentation defines these scopes:

claude mcp add uses local scope by default. So for an experiment, I would start there, use project scope for a server the team has agreed to share, and choose user scope when the same integration belongs in every repository I open.

Project configurations prompt for approval in interactive

Suppose you ask your coding assistant to find an existing price formatter before writing another one. Somewhere between that question and the answer, the application has to find a suitable tool, send it arguments, apply its execution controls, and put the result back where the model can use it.

MCP gives applications a shared protocol for that exchange. I want a result I can inspect: one that names a function, points me to its definition, and comes from a graph whose branch I know.

What an MCP server does

MCP is the Model Context Protocol, an open standard for connecting AI applications to external systems.

Think about maintaining integrations across several AI applications and several internal services. With a custom connector for each pairing, you have an N × M problem, where adding another application means revisiting every service connection, and changing a service can mean updating several connectors.

The shared interface lets compatible applications reuse a server integration.

Anthropic's MCP engineering article describes that integration problem.

The host, client and server

These names describe different responsibilities in the MCP architecture:

What does your agent need for the task you're about to give it?

For a library API question, I'd start with documentation. An error in production needs a different source, so I'd want the actual event, the stack trace, the project it came from, and the relevant release before asking for a fix. Those are different installs. I'd choose the missing source before choosing the server.

MCP means Model Context Protocol. It connects AI applications to external systems.

The picks below cover thirteen jobs from library docs to repository structure. I've linked the implementation or official plugin, included a setup command, marked the required credentials, and called out cases where a familiar server name now points at an archived project.

Before you paste a command

claude must be installed to run these examples. npx commands also need Node.js and npm.

For the client walkthrough, use MCP servers for Claude Code, with install commands.

Replace the uppercase credential and project placeholders, check the vendor's prerequisites, run the command, then open /mcp inside Claude Code to authenticate remote connections where required. The Claude Code MCP documentation covers registration and connection checks.

The Filesystem and Memory examples require Node.js 24 or