How Claude Code Connects to the Outside World: Models, Auth, MCP, Editors, and Remote Control

The adapter pattern means a team can switch model providers or add a new one without touching application logic—critical when costs, latency, or regional availability change. MCP and ACP turn the assistant into a composable component that plugs into existing editor workflows and internal toolchains, not a standalone app that demands its own silo.

The core agent loop only understands one internal event format. Model-provider adapters translate that common language into each vendor's proprietary API dialect—OpenAI, Gemini, or otherwise—and translate streaming responses, tool calls, and billing data back. Adding a new model means writing one adapter; the core, UI, and tool system never change. Authentication spans API keys, OAuth flows with automatic token refresh, and cloud-provider signatures, all unified behind a single auth state that the core loop never sees. Credentials land in the OS keychain first, falling back to encrypted files.

External tools, databases, and services connect through MCP, a socket standard that lets any compliant server offer tools, resources, and prompt templates to any compliant client. Configuration stacks across five levels—from company-wide admin policy down to per-project personal overrides—and external tools pass through the same permission gate as built-in ones. Editors drive the assistant via ACP, where the agent runs headless, reporting thinking streams, plan progress, and permission requests as structured messages; the editor owns all UI rendering. Remote control works the same way: the phone is a display and remote, the computer does the work, connected through a relay with QR-code pairing, encrypted channels, message throttling, and sleep prevention.