Best MCP Servers for Claude Code (2026): Scopes, Setup and a Context Budget

Most articles on this topic are a list of servers. The list is the easy part, and we already have one: our best MCP servers for coding roundup covers what each one does and which are worth having.

What that page deliberately does not cover is Claude Code's own machinery, and that machinery is where people actually get stuck: which scope to install into, what gets committed, and why adding the sixth server made everything slower.

Adding one

Three shapes, depending on how the server runs.¹

# Remote server over HTTP
claude mcp add --transport http notion https://mcp.notion.com/mcp

# Remote server that needs a header
claude mcp add --transport http secure-api https://api.example.com/mcp \
  --header "Authorization: Bearer your-token"

# Local server, launched by Claude Code itself
claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable \
  -- npx -y airtable-mcp-server

The -- in the third form matters: everything after it is the command Claude Code runs, not an option to claude mcp add.

If you already have servers configured in Claude Desktop, there is a shortcut rather than retyping them:

claude mcp add-from-claude-desktop

Scopes, which is the part people get wrong

Three scopes, and the default is not the one most people want.¹

claude mcp add --scope user <name> <url>      # everywhere, just you
claude mcp add --scope project <name> <url>   # this repo, the whole team

Use user for your own general tools. A documentation server or a browser server is something you want in every project, and installing it per project is how you end up configuring the same thing eleven times.

Use project for servers the repository needs. The internal API server that only makes sense inside one codebase belongs in that codebase, in .mcp.json, committed, so a new colleague gets it by cloning.

Use local almost never. It is the default, so it is what you get by accident: a server configured once, in one project, invisible from everywhere else, and forgotten by next week.

One rule about .mcp.json: commit the file, never the credentials. Pass secrets with --env or --header so the values live in your environment rather than in the repository.¹

The context budget nobody mentions

Here is the constraint that should drive your choices more than any feature comparison.

Every connected server's tool definitions are loaded into your context in every session, whether or not you use that server. Names, descriptions and parameter schemas for every tool it exposes. A server with twenty tools is twenty tool definitions in the window before you have typed anything.

Three consequences worth internalising:

1. Servers you rarely use are pure cost. They occupy the window in every session and pay you back in the one session a month you need them.
2. A server with many tools costs more than one with three. Tool count is a better proxy for context cost than how useful the server sounds.
3. More servers make tool selection worse, not just slower. The model is picking from a longer menu of similar-sounding options, and its accuracy at that falls as the menu grows.

If your sessions feel sluggish or your agent keeps reaching for the wrong tool, count your servers before you blame the model. Our piece on why Claude Code feels slow covers the other causes and the order to check them in.

A starting stack

Three servers, all at user scope, because these are things you want everywhere:

• A documentation server, so the agent stops inventing APIs.
• Your source host, so issues and pull requests happen without tab switching.
• A browser server, so the agent can check its own work rather than assert it.

That is it. Add a fourth when a real task has needed it twice, not when you read about one. The tool-by-tool argument for each of those, with the specific servers worth using, is in our MCP servers for coding roundup; if you want to build your own instead, we have how to build an MCP server.

Knowing what is actually loaded

Four commands, and it is worth running the first two now.¹

claude mcp list          # everything configured, and at which scope
claude mcp get <name>    # one server's details
claude mcp remove <name> # delete one
claude mcp login <name>  # sign in to a server that needs OAuth

Inside a session, /mcp shows live status and drives the browser OAuth flow for servers that require signing in.¹

Most people who run claude mcp list for the first time find at least one server they forgot they installed, which is the whole argument of this article in one command.

If a project's servers were approved and you want to be asked again:

claude mcp reset-project-choices

When a server is the wrong answer

Three cases, since this genre never mentions them:

• A one-off task. A script you run yourself is cheaper than a server you carry in every session afterwards.
• Something a file could do. Standing project knowledge belongs in CLAUDE.md, which costs a fraction of a server's tool definitions and cannot be misselected.
• A server you cannot describe in a sentence. If you cannot say when the agent should reach for it, neither can the agent.

FAQ

How do I add an MCP server to Claude Code? claude mcp add --transport http <name> <url> for a remote server, or claude mcp add <name> -- <command> [args] for a local one. Add --scope user for every project, or --scope project to share it through the repository.

What is the difference between the scopes? local is the default and covers the current project privately. project writes .mcp.json into the repository root and is shared through version control. user loads the server in all your projects. Local and user are stored in ~/.claude.json.

How many servers should I install? Fewer than you want to. Every connected server's tool definitions load into context in every session whether you use them or not, so start with three and add a fourth only when a real task has needed it twice.

How do I see what is loaded? claude mcp list from the shell, claude mcp get <name> for one server, and /mcp inside Claude Code for live status.

Should I commit .mcp.json? Yes when the servers are ones the whole team should have. Never commit credentials with it: pass them through --env or --header.

Can I use Claude Code itself as a server? Yes, claude mcp serve runs it as one.