Free Claude Code — No Telemetry, No Limits, All Hidden Features Unlocked
Free Claude Code — No Telemetry, No Limits, All Hidden Features UnlockedAnthropic forgot one line in a config file. Now everyone has their $2.5 billion source code.
This exists. Most people don’t know. Now you do.
512,000 lines of code. 44 secret features. A virtual pet nobody was supposed to see. And they shipped it all to a public website by accident. One security researcher typed one command. Found a 60MB file hiding in plain sight. Inside: everything that makes Claude Code tick — the AI coding tool that earns Anthropic $2.5 billion a year. Within hours, 41,500 people copied it. Someone rewrote it in Python before sunrise. Then Anthropic’s lawyers woke up. 8,100 copies got deleted. But the internet never forgets.
🧭 What's In This Post — Read This First
Not a developer? Don’t worry. This post explains everything like you’ve never touched a terminal.
| Section | What You’ll Learn |
|---|---|
| How it happened |
A $2.5B company forgot one line in a settings file |
 Hidden features |
20+ secret tools Anthropic built but didn’t release |
 Buddy pets |
A Tamagotchi living inside a coding tool (yes, really) |
 Undercover Mode |
AI that pretends to be human on GitHub |
 The free fork |
How someone stripped all tracking and unlocked everything |
 Security warnings |
What could go wrong (and a same-day hacking attack) |
 The DMCA war |
8,100 repos deleted — but the code lives on |
 What it phones home |
Every piece of data Claude Code sends to Anthropic |
 Next models leaked |
Opus 4.7 and a model that hallucinates 30% of the time |
 System prompts |
The actual instructions that tell Claude how to think |
 Best setup |
A hackathon winner’s 30-agent system, free to use |
 Free alternatives |
Tools that do the same thing — no subscription, no legal risk |
| Buddy species guide |
All 18 pets, rarity odds, and how to reroll yours |
 Secret settings |
Hidden switches that work on official Claude Code right now |
 Resources |
Every useful link in one table |
💣 How It Happened — One Line, $2.5 Billion Oops
Think of it like this: when a company builds software, they take the readable code and scramble it before shipping. They also create a “decoder ring” file (called a source map) that translates the scrambled version back to the original. This decoder ring is for internal use only. It should never leave the building.
Anthropic uses a tool called Bun to build Claude Code. Bun creates these decoder rings automatically. Someone on the release team forgot to tell the system “don’t include this file.” That’s it. One missing line in a settings file.
What actually shipped: A 59.8 MB decoder ring called cli.js.map — hidden inside the download package for Claude Code version 2.1.88.
What it decoded into: The complete, readable source code — 1,900 files, 512,000+ lines. Every tool. Every secret feature. Every internal project name. Everything.
| Question | Answer |
|---|---|
| Who found it? | A security researcher named Chaofan Shou |
| When? | March 31, 2026, around 4:23 AM |
| Anthropic’s excuse? | “Human error, not a security breach” |
| Were passwords or user data leaked? | No — just the tool’s code, not your stuff |
| Is this the first time? | No — they leaked a secret blog post 5 days earlier |
The funniest part: Inside the code, there’s a system called “Undercover Mode” — built specifically to stop internal secrets from leaking through AI-written code. They built a whole leak-prevention system… then leaked everything through a settings file.
👻 20+ Secret Features They Haven't Released Yet
The leaked code has 44 hidden switches that turn features on and off. Most are switched OFF in the version you download. Here’s what’s hiding:
| # | Feature | What It Does (Simple Version) |
|---|---|---|
| 1 | KAIROS | An AI that never sleeps — watches your project 24/7, writes daily notes, and does things without being asked |
| 2 | ULTRAPLAN | Sends your problem to a powerful computer in the cloud that thinks about it for up to 30 minutes |
| 3 | Coordinator Mode | One AI boss controlling multiple AI workers at the same time |
| 4 | Voice Mode | Talk to Claude with your actual voice instead of typing |
| 5 | Agent Triggers | Set Claude to automatically do tasks on a schedule (like an alarm clock for AI) |
| 6 | Daemon Mode | Claude runs invisibly in the background creating helper workers |
| 7 | Web Browser Tool | Claude can actually click around websites like a human (not just read them) |
| 8 | Buddy | A virtual pet that lives in your terminal — 18 species with stats like “Chaos” and “Snark” |
| 9 | autoDream | While you’re away, Claude reviews everything it learned and organizes its notes — like studying in its sleep |
| 10 | Peer Discovery | Claude finds OTHER Claude copies running on your computer and talks to them |
These aren’t ideas on a whiteboard. They’re finished code sitting behind on/off switches. Anthropic already built the future — they’re just releasing it slowly.
But here’s the catch: 108 pieces of the puzzle are missing. The most powerful internal tools were removed before shipping. You can see where they should plug in, but the rooms behind those doors are empty. The fork unlocks maybe 40% of what Anthropic employees actually use.
🐾 Buddy — Your Terminal Gets a Pet (This Is Real)
This officially launched on April 1, 2026. Type /buddy in Claude Code and a tiny ASCII creature hatches next to your cursor. It watches you code and comments in speech bubbles.
All 18 species (Anthropic hid these names using secret character codes):
duck · goose · blob · cat · dragon · octopus · owl · penguin · turtle · snail · ghost · axolotl · capybara · cactus · robot · rabbit · mushroom · chonk
How rare is yours?
| Rarity | Chance | What It Means |
|---|---|---|
| Common | 60% | Most people get this |
| Uncommon | 25% | A little special |
| Rare | 10% | Nice pull |
| Epic | 4% | Now we’re talking |
| Legendary | 1% | Flex material |
| Shiny (any rarity) | 1% extra | Rainbow shimmer effect |
| Shiny Legendary | 0.01% | ~1 in 10,000 people |
Every buddy has 5 stats: Debugging, Patience, Chaos, Wisdom, Snark. Plus hats — including a tiny rubber duck hat that only Legendary buddies can wear.
Your pet is tied to your account — same account always gets the same pet. But…
companionsection from~/.claude.json, restart Claude Code, and you’ll get a fresh roll. For full species control, the community tool any-buddy lets you pick exactly what you want.
Requires: Claude Code Pro subscription ($20/month minimum). Free users can’t access it.
See all 18 species: claude-buddy.vercel.app — interactive web gallery with every pet, hat, and stat combo.
Fake species warning: Names like “Nebulynx” and “Cosmoshale” circulating on social media are made up. Someone even launched a crypto coin ($Nebulynx on Solana) based on a species that doesn’t exist in the actual code.
🔍 Undercover Mode — AI That Pretends to Be Human
This is the most controversial finding. A file called undercover.ts (about 90 lines of code) tells Claude:
“You are operating UNDERCOVER in a PUBLIC/OPEN-SOURCE repository. Do not blow your cover.”
What it does in plain English:
- When Anthropic employees use Claude to write code for public projects, it removes every trace that an AI was involved
- Erases “Co-Authored-By: Claude” from commits
- Writes messages that sound like a human developer wrote them
- Tracks exactly what percentage of the code was AI-written… then hides that number
Who it affects: Only Anthropic employees. Regular users never see it — it’s removed before the tool reaches you.
Why it matters: If you’ve ever reviewed code from an Anthropic employee on GitHub — some of it was probably written by Claude. And you’d have no way to know.
⚡ The Free Fork — What It Actually Gives You (Honest Version)
A developer named Paolo took the leaked source and built a working copy called free-code. Here’s what changed:
Removed:
- All tracking (usage analytics, error reports, session fingerprinting)
- Content restrictions (hardcoded refusal patterns in the system prompt)
- The hourly “phone home” to Anthropic’s servers
- Managed settings that let Anthropic change your tool’s behavior remotely
Turned on:
- 45+ hidden feature switches (everything that still works without the missing internal code)
Install it:
curl -fsSL https://raw.githubusercontent.com/paoloanzn/free-code/main/install.sh | bash
export ANTHROPIC_API_KEY="sk-ant-..."
free-code
The Reality Check Nobody Tells You
| What They Say | What’s Actually True |
|---|---|
| “All features unlocked” | About 40% — 108 core pieces are missing from the leaked code entirely |
| “No tracking” | The tool doesn’t spy on you, but every question you ask still goes to Anthropic’s AI servers |
| “Guardrails removed” | Also removes protection against real attacks that can steal your API key |
| “Free” | You still need to pay for an API key — $3 to $25 per million words, depending on the AI model |
🛡️ Before You Install Anything — Read This
A Separate Attack Happened the Same Day
On the exact same day as the Claude Code leak (March 31, 2026), a completely different attack hit npm — the website where developers download code packages. North Korean hackers took over a popular package called axios and injected spy software into it.
| Question | Answer |
|---|---|
| What happened? | Hackers published fake versions of a common tool with hidden malware inside |
| When? | March 31, 12:21 AM to 3:29 AM (UTC) |
| Who’s at risk? | Anyone who downloaded ANY npm package during that 3-hour window |
| Does Claude Code use axios? | No — the timing is a coincidence, not a connection |
Check right now:
grep -r "1.14.1\|0.30.4\|plain-crypto-js" package-lock.json yarn.lock bun.lockb
If that command shows results → your computer might be compromised. Change all your passwords and API keys immediately.
Known Bugs in Claude Code That Forks Make Worse
| Bug | What It Does | Fixed In |
|---|---|---|
| CVE-2026-21852 | Bad code in a project can steal your API key before Claude even asks permission | Version 2.0.65 |
| CVE-2025-59828 | Some code runs before you agree to trust a project | Version 2.0.x |
Why this matters for the fork: Free-code removes the security system that protects against these exact attacks. The “no restrictions” feeling is fun until a random GitHub project drains your API credits.
Don’t: Run any fork on code you didn’t write yourself. Don’t use bypassPermissions. Don’t turn off the command injection check.
Fake packages alert: Scammers have already created fake npm packages pretending to be the leaked code. One called leaked-claude-code promises “no limits, no censorship” — classic malware bait. Don’t install random packages just because they mention the leak.
⚖️ The DMCA War — 8,100 Repos Deleted, Code Still Lives
Anthropic’s lawyers filed takedown notices within hours. GitHub killed 8,100+ copies in one sweep.
Dead:
| Mirror | Status |
|---|---|
| nirholas/claude-code (original) + 8,100 forks |  Gone |
| emmarktech/claude-code | DMCA’d |
| chatgptprojects/clear-code | DMCA’d |
| 0xKarl-dev/claw-codes | DMCA’d |
| 777genius/claude-code-source-code | DMCA’d |
Still alive:
| Project | What It Is | Why It Survived |
|---|---|---|
| claw-code | Rewritten from scratch in Python — 80K+ stars, fastest repo to hit 50K ever | No original code — completely new |
| Kuberwastaken/claude-code | Rewritten in Rust + the best technical breakdown online | Different language = different work |
| free-code | The working fork with backup on IPFS (a system with no delete button) | Modified + decentralized backup |
| Gitlawb mirror | Full original on a platform that has no central authority | Can’t send a takedown notice to nobody |
📡 What Claude Code Sends Back to Anthropic
Every time you use official Claude Code, it talks to four different tracking systems:
| System | What Gets Sent | Can You Stop It? |
|---|---|---|
| OpenTelemetry | How many words you used, how long your session lasted, what it cost | Yes |
| GrowthBook | Which features are turned on for you, what test group you’re in | Sort of |
| Sentry | Crash reports — includes the name of the folder you’re working in | Yes |
| Statsig | Usage patterns, which AI model is performing best | Sort of |
What it does NOT send: Your actual code, your questions, your files. Just how long your questions are.
What it sends that might surprise you: Your email, the folder you’re working in, what commands you ran (which can contain passwords), and your computer type.
6+ remote kill switches let Anthropic change how your Claude Code works — without asking you. It checks for new orders every hour.
To turn off tracking without using a fork:
export CLAUDE_CODE_ENABLE_TELEMETRY=0
export DISABLE_ERROR_REPORTING=1
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
🔮 What's Coming Next — Opus 4.7 and a Model That Lies 30% of the Time
Five days before the code leak, Anthropic had a SEPARATE leak — a website error showed ~3,000 unpublished files including a blog post about a secret model called Mythos (internal name: Capybara).
What the leaks reveal together:
- A new AI model tier above Opus — described as “larger and more intelligent”
- Internal notes show Capybara version 8 makes things up 29-30% of the time — that’s actually worse than version 4’s 16.7% error rate
- Code references to Opus 4.7 and Sonnet 4.8 (not publicly announced)
- The blog draft called Mythos “far ahead of any other AI model in cyber capabilities”
- Cybersecurity teams get early access
Two leaks in five days from the company that markets itself as “the careful AI company.” With 80% of their revenue from big businesses and an IPO coming, this pattern matters.
🧠 The Instructions That Tell Claude How to Think
The leaked code includes every set of instructions (called “system prompts”) that tell Claude Code how to behave. Two developers extracted them all:
Leonxlnx/claude-code-system-prompts — all prompts in one place
Piebald-AI/claude-code-system-prompts — better maintained, tracks 138+ versions
Interesting things the prompts reveal:
- KAIROS wakes up on a timer and decides whether to do something or keep sleeping — like a guard doing rounds
- Auto Mode uses a two-step safety check before running tools by itself (the function is literally called
classifyYoloAction()) - Coordinator Mode follows a 4-step plan: research → build → check → report
- A swear word detector uses basic pattern matching to know when you’re frustrated — an AI company using the simplest possible method to detect emotion is peak comedy
- Anti-copying protection injects fake tool names into conversations — if a competitor records and copies Claude’s behavior, the fake tools corrupt their copy
🧰 The Best Claude Code Setup — From a Hackathon Winner
Completely separate from the leak — an Anthropic hackathon winner gave away their entire Claude Code setup for free:
everything-claude-code by Affaan M.
| What’s Inside | Count |
|---|---|
| AI Agents (plan, review, fix, audit) | 30+ |
| Skills (test-driven dev, save tokens, memory) | 136+ |
| Commands (/plan, /tdd, /security-scan) | 60+ |
| Security tests | 1,282 |
Works on Claude Code, Cursor, OpenCode, and more. 100% free and open source. Built over 10+ months of real work.
🌐 Free Alternatives — No Subscription, No Legal Risk
You don’t need the leaked fork. These tools do the same job legally and for free:
| Tool | Works With | Best For | Stars |
|---|---|---|---|
| OpenCode | 75+ AI providers (ChatGPT, Gemini, Llama, local models) | Using any AI you want, for free | 95K+ |
| Aider | Any AI through LiteLLM | People who love Git | Mature |
| Cline | Multiple providers | VS Code users | Mature |
| claw-code | Multi-provider (planned) | Learning how Claude works | 80K+ |
Your Situation → Your Best Option
| You are… | Use this | Why |
|---|---|---|
| Curious, non-technical | CCLeaks.com | Browse every hidden feature in your browser — no install |
| A developer who wants flexibility | OpenCode + Ollama | Free, works offline, 75+ models, zero legal risk |
| Already paying for Claude | Official Claude Code + everything-claude-code | Get 30 agents and 136 skills on top of what you have |
| Wanting to study the architecture | Kuberwastaken/claude-code | Best written breakdown, Rust port |
| Just wanting hidden settings | Keep reading → next section  |
Works on official Claude Code, no fork needed |
The real lesson from this leak: Claude Code’s power isn’t in the code — it’s in the AI model behind it. Anyone can read the source now. Nobody can copy the intelligence.
🔧 Hidden Settings That Work on Official Claude Code Right Now
No fork needed. Set these in your terminal before running claude:
| Setting | What It Does | Tested? |
|---|---|---|
DISABLE_AUTO_COMPACT |
Stops Claude from compressing your conversation (keeps full memory) |  Confirmed |
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC |
Kills ALL non-essential network calls (including GrowthBook feature flags) | Confirmed |
CLAUDE_CODE_ENABLE_TELEMETRY=0 |
Turns off usage tracking | Confirmed |
DISABLE_ERROR_REPORTING=1 |
Stops crash reports (which include your folder path) | Confirmed |
CLAUDE_CODE_BRIEF |
Shorter, more compact responses | Unverified |
CLAUDE_CODE_PLAN_V2_AGENT_COUNT |
Override how many parallel agents run in plan mode | Unverified |
Full cheat sheet (auto-updated daily): cc.storyfox.cz — every command, shortcut, config option, and env variable. Printable. Maintained by a community member, verified on Hacker News.
Full env var list from source: GitHub Gist by jedisct1 — categorized as Documented / Advanced / Internal / Hidden.
📚 Every Resource — One Table
| What | Where | Notes |
|---|---|---|
| Browse the leak interactively |
CCLeaks.com | Best visual explorer — cited by The Register |
| Daily-updated cheat sheet |
cc.storyfox.cz | Every command, shortcut, and setting |
| Annotated source code |
sanbuphy/claude-code-source-code | Deep analysis docs included |
| System prompts |
Piebald-AI/claude-code-system-prompts | 138+ versions tracked |
| Free fork (stripped) |
paoloanzn/free-code | Use with caution — read security section |
 Python rewrite |
instructkr/claw-code | 80K+ stars, DMCA-proof |
 Rust port + breakdown |
Kuberwastaken/claude-code | Best technical analysis |
 Research-buildable fork |
beita6969/claude-code | Toggle individual feature flags |
 30-agent setup |
affaan-m/everything-claude-code | Hackathon winner’s config |
| Free alternative (any AI) |
OpenCode | 95K+ stars, works with 75+ providers |
| Buddy species gallery |
claude-buddy.vercel.app | See all 18 pets online |
| Pick your buddy |
cpaczek/any-buddy | Choose exact species + rarity |
 All env variables |
jedisct1 gist | Categorized by visibility level |
Quick Hits
| Want | Do |
|---|---|
| Browse the leak |
→ CCLeaks.com (no install needed) |
| Stop tracking |
→ CLAUDE_CODE_ENABLE_TELEMETRY=0 + DISABLE_ERROR_REPORTING=1 + DISABLE_NONESSENTIAL_TRAFFIC=1 |
| Free alternative |
→ OpenCode (works with any AI, zero cost with Ollama) |
| Best config |
→ everything-claude-code (30 agents, 136 skills) |
| See your buddy |
→ Type /buddy in Claude Code (Pro subscription required) |
| Cheat sheet |
→ cc.storyfox.cz (updated daily) |
| Python version |
→ claw-code (80K+ stars, can’t be DMCA’d) |
They spent years building the most advanced AI coding tool in the world. Then shipped the decoder ring. Security through obscurity lasted about 3 months.
Major update just dropped — added Buddy pet guide (reroll trick + all 18 species), hidden settings that work on official Claude Code right now, CCLeaks interactive explorer, fake package warnings, situation→action table so you know exactly what to use, and a full resource mega-table with 13 verified links. Give it a fresh read from the top. 
!