Subsystems Guide

Detailed documentation of Claude Code’s major subsystems.

Table of Contents

• Bridge (IDE Integration)
• MCP (Model Context Protocol)
• Permission System
• Plugin System
• Skill System
• Task System
• Memory System
• Coordinator (Multi-Agent)
• Voice System
• Service Layer

Bridge (IDE Integration)

Location: src/bridge/

The bridge is a bidirectional communication layer connecting Claude Code’s CLI with IDE extensions (VS Code, JetBrains). It allows the CLI to run as a backend for IDE-based interfaces.

Architecture

┌──────────────────┐         ┌──────────────────────┐
│   IDE Extension  │◄───────►│   Bridge Layer       │
│  (VS Code, JB)   │  JWT    │  (src/bridge/)       │
│                  │  Auth   │                      │
│  - UI rendering  │         │  - Session mgmt      │
│  - File watching │         │  - Message routing    │
│  - Diff display  │         │  - Permission proxy   │
└──────────────────┘         └──────────┬───────────┘
                                        │
                                        ▼
                              ┌──────────────────────┐
                              │   Claude Code Core   │
                              │  (QueryEngine, Tools) │
                              └──────────────────────┘

Key Files

Feature Flag

The bridge is gated behind the BRIDGE_MODE feature flag and is stripped from non-IDE builds.

MCP (Model Context Protocol)

Location: src/services/mcp/

Claude Code acts as both an MCP client (consuming tools/resources from MCP servers) and can run as an MCP server (exposing its own tools via src/entrypoints/mcp.ts).

Client Features

• Tool discovery — Enumerates tools from connected MCP servers
• Resource browsing — Lists and reads MCP-exposed resources
• Dynamic tool loading — ToolSearchTool discovers tools at runtime
• Authentication — McpAuthTool handles MCP server auth flows
• Connectivity monitoring — useMcpConnectivityStatus hook tracks connection health

Server Mode

When launched via src/entrypoints/mcp.ts, Claude Code exposes its own tools and resources via the MCP protocol, allowing other AI agents to use Claude Code as a tool server.

Related Tools

Configuration

MCP servers are configured via /mcp command or settings files. The server approval flow lives in src/services/mcpServerApproval.tsx.

Permission System

Location: src/hooks/toolPermission/

Every tool invocation passes through a centralized permission check before execution.

Permission Modes

How It Works

1. Tool is invoked by the Query Engine
2. checkPermissions(input, context) is called on the tool
3. Permission handler checks against configured rules
4. If not auto-approved, user is prompted via terminal or IDE

Permission Rules

Rules use wildcard patterns to match tool invocations:

Bash(git *)           # Allow all git commands without prompt
Bash(npm test)        # Allow 'npm test' specifically
FileEdit(/src/*)      # Allow edits to anything under src/
FileRead(*)           # Allow reading any file

Key Files

Plugin System

Location: src/plugins/, src/services/plugins/

Claude Code supports installable plugins that can extend its capabilities.

Structure

Plugin Lifecycle

1. Discovery — Scans plugin directories and marketplace
2. Installation — Downloaded and registered (/plugin command)
3. Loading — Initialized at startup or on-demand
4. Execution — Plugins can contribute tools, commands, and prompts
5. Auto-update — usePluginAutoupdateNotification handles updates

Related Commands

Skill System

Location: src/skills/

Skills are reusable, named workflows that bundle prompts and tool configurations for specific tasks.

Structure

Bundled Skills (16)

Execution

Skills are invoked via the SkillTool or the /skills command. Users can also create custom skills.

Task System

Location: src/tasks/

Manages background and parallel work items — shell tasks, agent tasks, and teammate agents.

Task Types

Task Tools

Memory System

Location: src/memdir/

Claude Code’s persistent memory system, based on CLAUDE.md files.

Memory Hierarchy

Related

• /memory command for managing memories
• remember skill for persisting information
• useMemoryUsage hook for tracking memory size

Coordinator (Multi-Agent)

Location: src/coordinator/

Orchestrates multiple agents working in parallel on different aspects of a task.

How It Works

• coordinatorMode.ts manages the coordinator lifecycle
• TeamCreateTool and TeamDeleteTool manage agent teams
• SendMessageTool enables inter-agent communication
• AgentTool spawns sub-agents

Gated behind the COORDINATOR_MODE feature flag.

Voice System

Location: src/voice/

Voice input/output support for hands-free interaction.

Components

Gated behind the VOICE_MODE feature flag.

Service Layer

Location: src/services/

External integrations and shared services.

See Also

• Architecture — How subsystems connect in the core pipeline
• Tools Reference — Tools related to each subsystem
• Commands Reference — Commands for managing subsystems
• Exploration Guide — Finding subsystem source code