How to Enable Hidden Features

Disclaimer

All analysis on this site is based on a specific version of the Claude Code source at instructkr/claude-code. Claude Code itself is not open-source. All intellectual property and credit belong to Anthropic. This is independent source code analysis, not official documentation, and is not affiliated with or endorsed by Anthropic in any way. Enabling hidden features may violate terms of service or cause unexpected behavior. Proceed at your own risk.

The Four Layers of Defense

Based on our source code analysis, hidden features are protected by four layers of gating:

Layer	Mechanism	Bypassable?
Build-time	USER_TYPE === 'ant' + feature() flags — code is physically stripped from public builds via dead-code elimination	Requires rebuilding from source
Runtime	GrowthBook tengu_* flags — server-side checks against Anthropic’s servers	Requires local override (CLAUDE_INTERNAL_FC_OVERRIDES)
Auth	Anthropic OAuth tokens tied to employee accounts	Not bypassable
Infrastructure	CCR, remote triggers, push notifications depend on Anthropic’s internal backend	Not bypassable

Environment Variables Worth Noting

Some features are partially gated by environment variables:

CLAUDE_CODE_COORDINATOR_MODE=1          # Coordinator mode
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1  # Agent teams (+ --agent-teams flag)
CLAUDE_CODE_PROACTIVE=1                 # Proactive mode
CLAUDE_INTERNAL_FC_OVERRIDES='{...}'    # GrowthBook flag overrides (JSON)

Important

These env vars only work if the corresponding code paths exist in your build. In the public build, most feature code has been tree-shaken away — the env vars have nothing to activate.

Quick Feasibility Assessment

Feature	Feasibility	Reason
Coordinator Mode	Speculative — may partially work	Most self-contained; env var + source rebuild
Agent Teams	Speculative — may partially work	Env var + CLI flag + source rebuild
Auto Permission Mode	Speculative — may partially work	Client-side classifier
Kairos (local tasks)	Speculative — partial	Local daemon works, but no push/webhooks
Voice Mode	Unlikely	May depend on server-side audio pipeline
CCR / Remote Triggers	Not possible	Requires Anthropic’s backend infrastructure
Bridge / SSH Remote	Not possible	Requires Anthropic’s remote backend

Speculative Approach: Building from Source

Speculation

The following is our best guess based on reading the build system. We have NOT verified these steps. They may be incomplete, outdated, or simply wrong.

Based on the source code structure, a theoretical approach might look like:

# 1. Clone the source
git clone https://github.com/instructkr/claude-code
cd claude-code

# 2. Install Bun (the runtime/bundler used by Claude Code)
curl -fsSL https://bun.sh/install | bash

# 3. Install dependencies
bun install

The key file to look for is the build configuration — likely a build.ts, bundle.ts, or a section in package.json scripts. Somewhere in the build pipeline, there should be --define flags like:

// Hypothetical build config (speculative)
Bun.build({
  define: {
    'process.env.USER_TYPE': JSON.stringify('ant'),  // Change 'external' to 'ant'
    // feature() flags are likely defined here too
  },
})

You would need to:

1. Find the build entry point that sets USER_TYPE and the feature() defines
2. Change USER_TYPE from its default (likely 'external') to 'ant'
3. Enable specific feature flags: COORDINATOR_MODE: true, KAIROS: true, etc.
4. Rebuild with bun build or whatever the project’s build script is
5. Replace your installed Claude Code binary with the custom build

The GrowthBook Problem

Even with a custom build, runtime tengu_* flags query Anthropic’s GrowthBook server. The override env var might help:

# Override specific GrowthBook flags locally
export CLAUDE_INTERNAL_FC_OVERRIDES='{
  "tengu_amber_flint": true,
  "tengu_kairos_cron": true,
  "tengu_onyx_plover": { "enabled": true, "minHours": 24, "minSessions": 5 },
  "tengu_scratch": true,
  "tengu_auto_mode_config": { "enabled": "enabled" }
}'

But this only works if:

• The CLAUDE_INTERNAL_FC_OVERRIDES code path exists in your build (it’s behind USER_TYPE === 'ant')
• The feature code itself wasn’t tree-shaken away

The Most Promising Targets

If we had to guess which features are most likely to work with a source rebuild:

1. Coordinator Mode — self-contained, env var activated, no backend dependency
2. Agent Teams — env var + CLI flag, uses local tmux/iTerm
3. Auto Dream — local file-based, forked subagent runs locally
4. Auto Permission Mode — client-side classifier, no server dependency
5. Scratchpad — just a directory with permission bypass, trivial

What Definitely Won’t Work

No amount of source hacking can replicate:

• CCR cloud environments — Anthropic’s private cloud infrastructure
• Remote triggers API (/v1/code/triggers) — Anthropic’s backend
• Push notifications — Anthropic’s notification service
• Bridge remote sessions — Anthropic’s environments API
• Features behind API beta headers — server-side rejection

To Be Continued…

We’re continuing to explore the build system and will update this page with verified steps if we find a reliable approach. In the meantime, the real value is in understanding how these features are designed — these architectural patterns (coordinator/worker, cache-safe forking, background memory consolidation, intelligent permission classification) can inspire your own AI tooling.

---

Source

All analysis is based on: github.com/instructkr/claude-code