Claude Code MCP Docs

What it does

• Four documented ways to add a server — remote HTTP, remote SSE, local stdio, and remote WebSocket — plus imports from setup instructions written for another client, from a URL, from an npx/uvx/binary command, or from an mcpServers JSON block
• Scope model spelled out: local, project, and user scopes, their hierarchy and precedence, project server approvals, and workspace trust
• Client runtime behavior in production terms: dynamic tool updates, notification streams on the v2 runtime, automatic reconnection, mid-session drops, failed first connections and retries, push messages with channels, and automatic backgrounding of long tool calls
• OAuth and auth depth: authenticate from the command line, fixed OAuth callback port, pre-configured credentials, overriding metadata discovery, restricting scopes, dynamic headers — plus environment variable expansion in .mcp.json with unset-variable and empty-credential semantics

When to use it

The ground truth for client-side MCP configuration in Claude Code — before debugging a server connection, check scopes, transport, and the failure-retry semantics documented here.

Getting started

No package — configuration. Local stdio example, in .mcp.json at the project root:

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "my-mcp-server"],
      "env": {
        "API_KEY": "${MY_API_KEY}"
      }
    }
  }
}

Environment variables expand with the documented syntax; unset variables follow the documented default semantics.

Things to know

Scope choice is a team decision as much as a technical one — project scope shares servers with everyone on the repo (and needs approval/trust), user scope does not; pick deliberately rather than accepting the default.