Voice, Browser & Remote: Unreleased Superpowers

Beyond the core CLI experience, Claude Code contains fully built but unreleased systems for voice interaction, browser automation, remote execution, cloud compute, and deep IDE integration. These features represent the next generation of Claude Code’s capabilities.

Voice Mode

Voice Mode adds real-time speech input to Claude Code, turning the terminal into a voice-driven coding assistant.

How It Works

Real-time voice-to-text streaming via voiceStreamSTT.ts — speech is transcribed as you talk, not after you finish
Push-to-talk activation bound to the space bar keybinding
Key term boosting through voiceKeyterms.ts — technical vocabulary (API names, language keywords, framework terms) gets priority in transcription to reduce errors on developer jargon

Gate Structure

All three conditions must be true simultaneously:

feature('VOICE_MODE') feature flag is enabled
isVoiceGrowthBookEnabled() returns true (GrowthBook remote config)
isAnthropicAuthEnabled() confirms Anthropic authentication is active

The killswitch flag tengu_amber_quartz_disabled can shut down Voice Mode remotely without a code deploy.

Web Browser Tool

Claude Code includes a built-in browser automation system with two distinct approaches:

Native WebView (Bun API)

A browser automation layer built on top of Bun’s WebView API. Claude can:

Navigate to URLs
Click elements and fill forms
Take screenshots of pages
Evaluate arbitrary JavaScript

Chrome DevTools Integration (`claudeInChrome`)

An MCP server that connects directly to Chrome DevTools Protocol, providing deeper browser control:

Full page navigation and interaction
Form filling and element selection
Screenshot capture
Script evaluation in page context

This dual approach gives Claude both a lightweight built-in browser and the ability to control a full Chrome instance when needed.

SSH Remote Mode

SSH Remote Mode allows Claude Code to execute commands on remote machines over SSH, extending the agent beyond the local environment.

Key Components

RemoteSessionManager.ts — manages SSH connections and session lifecycle
SessionsWebSocket.ts — WebSocket transport layer for real-time communication with remote hosts

Remote Permission Bridge

When Claude needs to perform a privileged action on a remote machine, the permission system works across the SSH boundary:

Claude requests permission for a remote action
The request is forwarded back to the local session
The user approves or denies locally
The decision is relayed to the remote execution context

This ensures you never lose control over what Claude does on remote machines.

Bridge System / IDE Integration

The Bridge System is Claude Code’s IDE integration protocol, implemented across 30+ files in src/bridge/. This is the backbone that powers the VS Code extension and other IDE integrations.

Multi-Session Support

A single environment can host up to 32 concurrent Claude sessions, each independently managed. This enables parallel workstreams within one IDE workspace.

Spawn Modes

Mode	Behavior
`single-session`	One Claude session at a time
`worktree`	Each session gets an isolated git worktree — full filesystem isolation
`same-dir`	Multiple sessions share the same working directory

Session Lifecycle

register  -->  poll  -->  spawn  -->  heartbeat  -->  done

Register — IDE registers a new session request
Poll — Claude polls for pending session requests
Spawn — Session is created and Claude begins work
Heartbeat — Periodic keepalive between IDE and Claude
Done — Session completes and resources are cleaned up

Permission Flow

The Bridge System implements a full permission delegation chain:

Claude session requests permission for an action
IDE receives the request via WebSocket
User decides (approve/deny) within the IDE UI
Response is forwarded back to the Claude session

Security

JWT authentication for session identity verification
Trusted device verification to prevent unauthorized connections
Reconnection support — sessions can resume after a bridge crash without losing state

CCR (Cloud Compute Resource)

CCR enables Claude Code to offload heavy computation to Anthropic’s cloud infrastructure. Remote agent tasks run server-side with full access to Claude’s capabilities.

Remote Task Types

Task Type	Purpose
`ultraplan`	Multi-phase planning with structured output
`ultrareview`	Deep code review with verification
`autofix-pr`	Automated PR fixes
`background-pr`	Background pull request creation

UltraPlan

A remote multi-phase planning system:

needs_input phase — Claude asks clarifying questions before planning
plan_ready phase — the finalized plan is delivered back to the local session

This offloads expensive planning work to cloud compute where longer context windows and more compute are available.

UltraReview

A remote code review system that goes beyond surface-level linting:

Bug finding — deep analysis for logical errors and edge cases
Verification — automated checks against the identified issues
Refutation tracking — tracks when a flagged issue is actually a false positive

Undercover Mode

Undercover Mode is a leak-prevention system designed to keep internal details out of public repositories.

Behavior

Auto-detects when the user is working in a public or open-source repository
Strips sensitive content from commits and PRs:
- Internal model codenames
- Internal project names
- Slack channel references
Always ON unless the repository is explicitly in the internal allowlist (INTERNAL_MODEL_REPOS)
No force-OFF option — this is intentional to prevent accidental leaks

Environment Variable

CLAUDE_CODE_UNDERCOVER=1

This env var forces Undercover Mode ON regardless of user type, useful for testing or for extra caution in sensitive contexts. There is no equivalent variable to force it off.