QuetzaCodetl: The Uncensored Claude Code
A local-first fork with zero content restrictions. No API keys, no logging to Anthropic, no filters — just your code on your hardware.
Why We Built an Uncensored Claude Code
Claude Code is an incredible tool. But it has a problem: Anthropic decides what you can and cannot code. Every prompt runs through their content filters. Every session is logged on their servers. Every refusal is a black box — no explanation, no appeal, no override.
If you're building an adult platform, writing a security tool, researching AI alignment, or just don't want your entire development workflow logged by a corporation — Claude Code is not your tool.
QuetzaCodetl is the answer. It's a local-first coding assistant that runs entirely on your hardware. The same interface, the same tool system, the same workflow — but zero content restrictions and zero external logging.
Claude Code vs QuetzaCodetl
| Feature | Claude Code | QuetzaCodetl |
|---|---|---|
| Content restrictions | Anthropic's filters block NSFW, security research, and "harmful" code | None — abliterated models, no guardrails |
| Privacy | Every prompt logged on Anthropic servers | 100% local — nothing leaves your machine |
| API keys required | Anthropic API key + billing | None — runs on llama.cpp, completely free |
| Model choice | Only Claude models, only via API | Any GGUF model — Qwen, Gemma, DeepSeek, GLM, open-source |
| Offline use | Requires internet connection | Fully offline — works air-gapped |
| Cost | Per-token pricing, adds up fast | Free — uses your GPU, no usage limits |
| MCP tools | Full MCP support | Full MCP support — Odin search, desktop agent, browser automation |
| Context window | 200K (Claude) | Up to 1M (depending on model) |
How It Works
QuetzaCodetl is a local-first coding assistant built on the Claude Code tool-use architecture — but instead of calling api.anthropic.com, it runs entirely on your own hardware with whatever model you choose.
The key insight: Claude Code's value isn't Claude — it's the tool-use architecture. The agent loop (think → call tools → observe → think) works just as well with local models. With abliterated models — where the refusal vectors have been surgically removed from the weights — you get the full coding agent experience with no refusals, ever.
We run it on a single RTX 4090 24GB. With IQ2 quantization, models like Qwen3-Coder-Next hit 86 tok/s at 200K context — faster than Claude Code's API latency, and completely free.
What "Uncensored" Actually Means
Let's be precise. "Uncensored" doesn't mean "malicious." It means:
No content filter says no. Claude Code will refuse to help with penetration testing tools. It'll refuse NSFW content generation. It'll refuse certain security research topics. It'll even refuse to discuss the refusal — you just get stonewalled with a vague "I can't help with that." QuetzaCodetl will help you write whatever code you want, period. You're an adult. You decide what's appropriate.
No logging to third parties. Anthropic stores every conversation. Your code, your project architecture, your internal tools — all sitting on someone else's servers. QuetzaCodetl stores conversations locally. Nothing leaves your machine unless you explicitly set up an API-backed model.
No rate limits, no quotas. Your GPU, your rules. Run it 24/7. Batch process hundreds of files. No one is counting tokens and sending you a bill.
Models We Recommend
QuetzaCodetl works with any GGUF model. These are the ones we use daily:
Qwen3-Coder-Next (abliterated) — IQ2_XS, 86 tok/s
Our daily driver. The abliterated version has refusal vectors removed. It handles NSFW code, security tools, and anything Anthropic would block. 200K context at blazing speed.
Gemma 4 26B MoE — Q4_K_M, ~40 tok/s
Google's open model with Mixture of Experts architecture. Strong reasoning, good for architecture planning. Not abliterated but has minimal refusals compared to Claude.
GLM-4.7 Flash — via API, 1M context
When you need the million-token context window. Good for massive codebase analysis. API-based so not fully local, but no Anthropic-style content filtering.
DeepSeek V4 — via API, 1M context
The frontier option. DeepSeek's models have far fewer content restrictions than Anthropic's. Great for complex multi-file refactors where you need the full context.
The MCP Ecosystem
QuetzaCodetl isn't just a code editor — it's a full agent platform with MCP (Model Context Protocol) tools that Claude Code users will recognize:
Odin Search — Self-hosted web search engine. No Google tracking. Map the internet's authoritative hubs and fetch live content. 250K+ pages indexed.
Desktop Agent — Screen understanding via AT-SPI + OCR. Give a blind local LLM the ability to see and interact with your desktop. Click buttons, fill forms, automate any app.
Agent Browser — Headless browser automation with authenticated Chrome profiles. Scrape, screenshot, and interact with any website.
Letta Memory — Persistent cross-session memory. The agent remembers your preferences, past decisions, and project context across sessions.
The Bottom Line
If you're happy with Claude Code and never hit content refusals, there's no urgent reason to switch. But if you've ever been blocked mid-workflow because Anthropic decided your code was "inappropriate" — or if you simply don't want your entire development process logged on someone else's servers — QuetzaCodetl is the answer.
It's the same tool-use architecture. The same MCP ecosystem. The same workflow. No filters. No logging. No keys. No bills.
Interested in running an uncensored Claude Code alternative? Read our guide to running uncensored AI locally or get in touch.