Hidden Features
The Claude Code source contains several feature flags and capabilities that are not publicly documented. These were found by reading the source code directly.
Feature flags
Autonomous mode. Claude receives <tick> prompts that keep it alive and acts without user input. When terminal is unfocused it commits, pushes, and explores independently.
Multi-agent orchestration. The main session acts as a coordinator that dispatches work to sub-agents and synthesizes their results.
IDE integration bridge. Connects Claude Code to VS Code, JetBrains, and other editors for inline suggestions and context sharing.
Background daemon mode. Claude runs as a persistent background process, processing tasks without a visible terminal session.
Voice input support. Enables speech-to-text for submitting prompts. Infrastructure exists in the source but not publicly released.
Enables CronCreate, CronDelete, CronList and RemoteTrigger tools. Allows scheduling recurring agent tasks on a cron schedule.
Enables the MonitorTool for performance and observability monitoring during agent execution.
Auto-cleanup of old tool results in the context window. Replaces stale Read/Bash/Grep outputs with [Old tool result content cleared] to free token budget.
Skill discovery mechanism. Dynamically finds and loads available skills from the plugin system based on the current task context.
Activates the built-in Verification Agent that runs in the background after implementations and emits PASS / FAIL / PARTIAL verdicts.
Hidden capabilities
Beyond feature flags, the source reveals several built-in capabilities that are not prominently documented.
Undercover mode
For internal Anthropic users, model names and IDs are hidden from the system prompt. This prevents unannounced models from leaking into commits, PRs, or conversation history.
Scratchpad
A temporary, isolated directory created per session. Claude can write freely here without triggering permission prompts. It's an alternative to /tmp and is cleaned up after the session.
User hooks
Shell commands configured in settings.json that execute in response to events (tool calls, session start, etc.). Claude treats hook output as if it came directly from the user.
Worktree detection
Claude detects if it's running inside a git worktree and adjusts behavior accordingly. Specifically, it will not cd to the original repo root and respects the worktree isolation.
settings.json and Claude treats their output as
user messages. This makes hooks a powerful way to automate workflows without changing
your prompts.