MCP servers

The Model Context Protocol connects Claude to external services: databases, APIs, browsers, issue trackers. Servers come in three transport types: stdio (local process), SSE (hosted, often OAuth), and HTTP (REST endpoints).

Team-shared servers live in .mcp.json at the project root; personal servers go in ~/.claude.json (add with claude mcp add --scope user). Precedence when names collide: local > project > user.

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "${GITHUB_TOKEN}"
      }
    },
    "docs-api": {
      "type": "http",
      "url": "https://api.example.com/mcp",
      "headers": {
        "Authorization": "Bearer ${DOCS_API_KEY}"
      }
    }
  }
}

The ${GITHUB_TOKEN} reference is read from your shell environment when the server starts — the token never lands in the file. MCP tool names follow mcp__server__tool, which is also how you target them in hook matchers and permission rules.

Context cost is low by default: tool search defers full schemas until Claude actually uses a tool; only names load at session start. Run /mcp for connection status and per-server token costs.

Check your understanding

Question 1Where do team-shared MCP servers belong?

Question 2Why write ${GITHUB_TOKEN} instead of the token itself?

Question 3What keeps idle MCP servers cheap in context?