MCP in Claude Code: how to add and use MCP servers

On its own, Claude Code can read and edit files, run commands and fetch web pages. The Model Context ProtocolModel Context ProtocolAn open standard for connecting AI applications to tools and data. A system offered once as an MCP server works in any application that speaks MCP. (MCP) is how you give it more: an MCP server is a small program that offers Claude a set of tools for one system, such as your issue tracker, a database, Figma or Sentry. Once it is connected, Claude reads the ticket or runs the query itself instead of working from what you paste in.

A good sign that you need one: you keep copying text from another tool into the chat.

Add a remote server

Most services run their MCP server on the web. Add it by name and address:

claude mcp add --transport http notion https://mcp.notion.com/mcp

Then start Claude Code, run /mcp, choose the server and sign in. Many remote servers use OAuth, so you sign in through your browser once and Claude Code keeps the token. Servers that take a token instead get it as a header:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
  --header "Authorization: Bearer $GITHUB_TOKEN"

Use a token with access to only the repositories Claude should work in. claude mcp add saves the settings without testing them, so check the server shows as connected in /mcp.

Add a server that runs on your machine

Some servers are a program you run locally. Put the command that starts it after --:

claude mcp add --transport stdio db -- npx -y @bytebase/dbhub \
  --dsn "postgresql://readonly:secret@localhost:5432/shop"

Everything after -- is passed to the server untouched; everything before it is for Claude Code. Use --env KEY=value to give the server a variable. The example connects with a read-only database user, so no query Claude writes can change data.

Then ask in plain words: "show me the schema of the orders table", "how many orders were refunded last week?"

Where servers are saved: 3 scopes

Use project scope for servers the whole team needs, and commit .mcp.json. Never put a secret in it: write ${VAR} and each person's environment fills it in, with ${VAR:-fallback} for a default:

{
  "mcpServers": {
    "sentry": {
      "type": "http",
      "url": "https://mcp.sentry.dev/mcp"
    },
    "internal-api": {
      "type": "http",
      "url": "${API_BASE_URL:-https://api.example.com}/mcp",
      "headers": { "Authorization": "Bearer ${INTERNAL_API_TOKEN}" }
    }
  }
}

Claude Code asks each person before it uses a server from .mcp.json, since anyone with commit access could have added it.

Managing servers

claude mcp list             # every server and whether it is connected
claude mcp get github       # one server's settings
claude mcp remove github    # remove it, and its saved sign-in

Inside a session, /mcp shows each server's status and tools, and lets you sign in again or turn a server off without removing it.

Using what a server offers

• Tools. Ask for what you want and Claude calls the right tool. In Manual mode, each call asks permission like any other action until you allow it; mcp__github in your allow rules covers every tool of the github server.
• Resources. Type @ to see a server's resources next to your files, and mention one, such as @github:issue://123, to put it into your message.
• Prompts. A server's prompts appear as commands: type / and look for the ones marked (MCP).

Context and output size

Every connected server adds tools Claude could call. Claude Code uses tool search for them: only the tool names load at the start, and a tool's full description loads when Claude needs it. So adding servers costs little context until their tools are used. /context shows what they take.

A tool that returns a lot can still fill the conversation. Claude Code warns when one result passes 10,000 tokens. A result over the limit, 25,000 tokens by default, is saved to a file that Claude reads when it needs to. Raise the limit with the MAX_MCP_OUTPUT_TOKENS environment variable.

Check a server before you trust it

An MCP server runs with whatever access you give it. Before you connect one:

• Prefer servers from the service's own maker or a reviewed directory.
• Give it the narrowest token or database user that does the job.
• Be careful with servers that bring in outside text, such as web pages, emails or support tickets. That text can carry instructions written for Claude, which is Prompt InjectionPrompt InjectionAn attack that hides instructions in text a model reads, such as a web page, an email or a file, so the model follows the attacker instead of you..
• Keep an ask rule on tools that change things. /mcp lists each server's tools by name, and a rule such as mcp__github__<tool> names one.