How to Run Uncensored GLM-4.7-Flash Locally with Claude Code

The Problem with Commercial AI

You're working on a creative project. Maybe it's a novel with mature themes, or research into a controversial historical period, or code for a sensitive application. You ask your AI assistant for help, and you get that response:

"I cannot help with that."

Not because the request is illegal. Not because it's harmful. But because a corporate policy decided this topic is off-limits—even for legitimate use cases.

This happens daily to developers, writers, researchers, and creators. The restrictions aren't about safety; they're about control.

I got tired of it. So I built something different.

Introducing GLM-4.7-Flash-Rapport

GLM-4.7-Flash-Rapport is a complete system for running uncensored AI locally, integrated with Claude Code. It combines:

• GLM-4.7-Flash-PRISM: An ablitterated (uncensored) open-weights model
• 198k context window: Massive conversation memory
• Google Search: Real-time web access via API
• Vision capabilities: Image analysis with smart routing
• Local execution: Complete privacy, zero subscriptions

The result? An AI that helps you without arbitrary restrictions.

What "Ablitterated" Actually Means

The term comes from AI research circles. Most commercial models are trained with refusal mechanisms—they learn to say "I cannot help" when certain patterns are detected.

Ablitteration removes these refusal behaviors while keeping the model's actual capabilities intact. The AI doesn't lose intelligence; it gains honesty.

Think of it like removing training wheels. The model can now respond directly to your requests, trusting you to use it responsibly.

What You'll Need

Hardware Requirements

• NVIDIA GPU with 24GB VRAM (tested on RTX 4090)
• Linux system (Ubuntu 22.04+ recommended)
• 32GB+ system RAM
• 50GB storage

Don't have the hardware? You can use cloud GPU rental services like RunPod or Vast.ai.

Skills Required

• Basic Linux command line knowledge
• Comfort with environment variables
• Ability to obtain and configure API keys
• Patience for troubleshooting (it's worth it)

Step-by-Step Installation

Step 1: Install Prerequisites

First, get llama.cpp with CUDA support:

Install Node.js and Python packages:

Step 2: Download the Model

Step 3: Install the Wrapper

Step 4: Configure Environment Variables

Create ~/.glm-flash-env:

Load it: source ~/.glm-flash-env

Step 5: Get API Keys

OpenRouter (for vision):

1. Visit openrouter.ai
2. Create account → Generate API key

Google Custom Search (for web search):

1. Go to Google Cloud Console
2. Create project → Enable "Custom Search API"
3. Create API credentials → Copy key
4. Go to Google Custom Search
5. Create search engine → Enable "Search the entire web"
6. Copy the Search Engine ID (CX)

Pro tip: The "Search the entire web" toggle is crucial—without it, you only search specific sites.

Step 6: Configure MCP Servers

The easiest way:

This automatically configures Claude Code to use your MCP servers.

Step 7: Install Skills and Link

Step 8: Test It

How It Works Under the Hood

The architecture is elegant:

1. Claude Code sends requests to the wrapper
2. Wrapper detects content type (text vs images)
3. Text → routes to local GLM model via llama.cpp
4. Images → routes to OpenRouter vision API
5. Google Search → uses MCP server for web queries
6. Everything stays in your conversation history locally

Image Routing Architecture

Web Search Integration

Key insight: The system is modular. Don't want Google Search? Remove the MCP server. Want different vision? Swap the OpenRouter model. It's your setup.

Real-World Use Cases

For Writers

Create fiction with mature themes without content warnings. Research sensitive historical periods without sanitization. Develop diverse characters without artificial constraints.

For Developers

Generate code for any domain. Debug proprietary systems. Build tools without restriction on functionality.

For Researchers

Analyze controversial topics objectively. Access scientific discussions without corporate bias. Build knowledge bases from diverse sources.

For Artists

Create any visual concept. Explore boundaries without platform restrictions. Maintain ownership of your creative process.

Troubleshooting

"Server won't start"

• Check logs: tail -f /tmp/glm-flash-server.log
• Verify model path exists
• Test llama.cpp directly: ./llama-cli -m model.gguf -p "test"

"Google Search not working"

• Most common cause: "Search the entire web" not enabled
• Test API: curl "https://www.googleapis.com/customsearch/v1?key=$GOOGLE_SEARCH_API_KEY&cx=$GOOGLE_SEARCH_CX&q=test"

"Vision not working"

• Check OpenRouter key is set
• Verify proxy is running: ps aux | grep image-routing
• Check proxy logs: tail -f /tmp/glm-flash-image-routing-proxy.log

The Philosophy

This isn't about bypassing safety. It's about user agency.

Commercial AI services make decisions for you about what's appropriate. They decide which topics are allowed, which code is acceptable, which creative directions are permitted.

Running local AI shifts that responsibility to you—where it belongs.

If you're writing a novel with dark themes, you should be able to explore those themes. If you're researching a controversial historical event, you should get unfiltered information. If you're building software, you should be able to generate any code you need.

The AI is a tool. You decide how to use it.

Is It Worth the Effort?

Honestly? For some people, no.

If you're happy with ChatGPT's restrictions, if you don't mind paying subscriptions, if you trust corporate AI policies—stick with what's working.

But if you've ever:

• Hit an artificial wall with "I cannot help"
• Wanted privacy for sensitive work
• Needed an AI that actually does what you ask
• Been frustrated by policy changes that break your workflow

Then yes, it's absolutely worth it.

What's Next?

The world of local AI is evolving fast. Projects like Ollama, LM Studio, and countless open-source models are making this accessible to everyone.

This setup is just one configuration. You could:

• Add file system tools
• Integrate with databases
• Build custom MCP servers
• Connect to your entire development environment

The only limit is what you want to build.

Resources

• GitHub: github.com/Indras-Mirror/GLM-4.7-Flash-Rapport
• Full Guide: indrasmirror.au/blog-running-uncensored-ai-local.html
• Need Help?: AI Setup Consultation

Final Thoughts

We're at an inflection point with AI. The technology is becoming ubiquitous, but control is concentrating in fewer hands.

Running your own AI isn't just about privacy or avoiding subscriptions. It's a statement that you deserve to own your tools.

The setup is technical. It requires time and effort. But the result is something increasingly rare: a piece of technology that works for you, not for shareholders.

Give it a try. Your future self will thank you.