GLM-5 MCP Server for Claude Desktop

Reduce Claude Desktop consumption by 10x by delegating heavy tasks to Z.ai's GLM-5 (744B parameter) model through the Model Context Protocol (MCP).

🎯 Problem This Solves

Are you hitting Claude Pro limits too fast?

• Weekly limit exhausted in 2 days? ✅
• Blocked from all models for days? ✅
• Paying $100/month but can't use it 5 days/week? ✅

This MCP server gives you:

• 10x reduction in Claude consumption (Opus 4.6 → Sonnet 4.5 + GLM-5 delegation)
• 5x reduction in Claude consumption (Opus 4.6 → Opus 4.6 + GLM-5 delegation)
• Continuous availability - never blocked again
• Cost-effective scaling - $40-60/month Z.ai vs. $100 to 200$ /month paid additional to claude to continue using
• 18-30x ROI on your Claude Pro subscription

🚀 Quick Start

1. Get Z.ai API Key

1. Visit Z.ai and create an account
2. Navigate to API Keys
3. Create a new API key
4. Copy your key (format: xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx.xxxxxxxxxxxxxxxx)

2. Install

# Clone the repository
git clone https://github.com/Arkya-AI/glm5-mcp.git
cd glm5-mcp

# Install dependencies
npm install

3. Configure Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "glm5": {
      "command": "node",
      "args": [
        "/ABSOLUTE/PATH/TO/glm5-mcp/index.js"
      ],
      "env": {
        "ZAI_API_KEY": "your-z-ai-api-key-here"
      }
    }
  }
}

Important: Replace /ABSOLUTE/PATH/TO/ with the actual path to where you cloned the repo.

4. Restart Claude Desktop

Quit Claude Desktop completely and restart it. The GLM-5 tools will now be available.

🛠️ Available Tools

Core AI Tools

ask_glm5

Delegate complex reasoning tasks to GLM-5 (744B parameters).

Use for:

• Complex analysis and reasoning
• System design and architecture
• Advanced problem-solving
• Multi-step logical tasks

Parameters:

• prompt (required): Your task/question
• system_prompt (optional): Custom behavior guidance
• temperature (optional): 0.0-1.0, default 0.7
• max_tokens (optional): Max response length, default 4000

ask_glm5_pro

Same as ask_glm5 but with coding-optimized system prompt.

Use for:

• Code generation
• Refactoring and optimization
• Debugging assistance
• Technical implementation

Research & Intelligence Tools

web_search

LLM-optimized web search powered by Z.ai.

Use for:

• Competitive intelligence
• Market research
• Real-time news and trends
• Finding multiple sources

Parameters:

• search_query (required): Your search query
• count (optional): Results 1-50, default 10
• search_recency_filter (optional): oneDay, oneWeek, oneMonth, oneYear, noLimit
• search_domain_filter (optional): Comma-separated domains

web_reader

Fetch and parse full web page content.

Use for:

• Reading articles and blog posts
• Analyzing competitor pages
• Extracting documentation
• Deep content analysis

Parameters:

• url (required): URL to fetch
• return_format (optional): markdown or text, default markdown
• with_images_summary (optional): Include image summary, default false
• with_links_summary (optional): Include links summary, default false
• timeout (optional): Timeout in seconds, default 20

parse_document

Extract text from PDFs and images using GLM-OCR.

Use for:

• PDF proposals and contracts
• Scanned documents
• Business cards
• Invoices and receipts
• Complex layouts and tables

Parameters:

• file_url (required): Public URL to document/image
• return_format (optional): markdown or text, default markdown
• parse_mode (optional): auto, ocr, or layout, default auto

Supports: PDF, images up to 50MB or 100 pages

📊 Usage Strategy

Execution Priority (follow this order)

When Claude faces a task, it should follow this priority chain:

Sub-Agent Enforcement

Task sub-agents spawned by Claude MUST also use GLM-5 for their heavy work. When spawning a sub-agent, explicitly instruct it to use ask_glm5_pro for code generation and ask_glm5 for analysis/docs. Sub-agents that generate >50 lines of code or >300 words of content themselves (without delegating to GLM-5) are violating the delegation model.

Pattern: GLM-5 generates → Claude/sub-agent writes to disk.

Orchestration Model

Claude / Opus (Parent)
    ├── Planning & coordination (stays in Claude)
    ├── File operations & disk I/O (stays in Claude)
    ├── Quick responses <100 words (stays in Claude)
    │
    ├── PRIORITY 1: Spawn parallel sub-agents (for multi-part tasks)
    │    └── Each sub-agent uses GLM-5 for code/analysis
    │         ├── ask_glm5_pro for code generation
    │         ├── ask_glm5 for analysis/docs
    │         └── Sub-agent writes output to disk
    │
    └── PRIORITY 2: Delegate to GLM-5 directly (for single-unit tasks)
         ├── Analysis >300 words
         ├── Code generation >50 lines
         ├── Research synthesis from multiple sources
         └── Document processing & OCR

Model-Specific Strategies

Sonnet 4.5 (default for most tasks):

• Orchestration, file ops, disk I/O, quick responses
• Delegate all analysis >300 words and code >50 lines to GLM-5
• Result: 10x token reduction

Opus 4.6 (complex multi-step coordination):

• Priority 1: Spawn parallel sub-agents (each uses GLM-5)
• Priority 2: Delegate to GLM-5 directly
• Priority 3 (last resort): Opus does it itself
• Pattern: Opus = Architect, Sub-agents = Parallel workers, GLM-5 = Content engine
• Result: 90%+ reduction in Opus consumption

Example Workflows

Research pipeline:

1. Start session with Sonnet 4.5
2. Use web_search to find sources
3. Use web_reader to fetch content (parallel)
4. Use ask_glm5 to analyze and synthesize
5. Sonnet formats and presents results

Multi-part task (Opus):

1. Opus analyzes task and identifies independent parts
2. Spawns parallel sub-agents for each part
3. Each sub-agent calls ask_glm5 or ask_glm5_pro
4. Each sub-agent writes output to disk
5. Opus integrates results

Expected Savings

Before GLM-5 MCP:

• Weekly limit in 2 days
• 5 days blocked per week
• $100/month for 28% availability

After GLM-5 MCP:

• Sonnet 4.5: 10x less quota usage
• GLM-5: Handle heavy lifting
• 100% continuous availability
• $140-160/month total cost

ROI: 18-30x improvement in effective cost per hour of usage

🔧 Development

Project Structure

glm5-mcp/
├── index.js           # Main MCP server
├── package.json       # Dependencies
├── README.md          # This file
├── LICENSE            # MIT License
└── .gitignore         # Git ignore rules

Testing

# Test the MCP server
npm start

The server will start in stdio mode and log to stderr. Use Claude Desktop to test the tools.

Adding New Tools

1. Add tool definition to ListToolsRequestSchema handler
2. Add tool handler in CallToolRequestSchema handler
3. Update README documentation
4. Test in Claude Desktop

📖 Documentation

For Users

• README.md - Setup and installation guide (you're here!)
• EXAMPLES.md - Real-world usage examples
• CONTRIBUTING.md - How to contribute

For Claude

• CLAUDE.md - Comprehensive project memory and delegation guidelines
  • When to use each tool (Sonnet 4.5 vs Opus 4.6 strategies)
  • Code quality standards and patterns
  • Optimization rules to reduce Claude consumption
  • Development workflow and troubleshooting

API Reference

This MCP server uses the Z.ai API. Key endpoints:

• /paas/v4/chat/completions - GLM-5 text generation
• /paas/v4/web_search - Web search
• /paas/v4/reader - Web content fetching

See Z.ai Documentation for complete API reference.

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Development Guidelines

1. Follow existing code style
2. Add tests for new features
3. Update documentation
4. Keep tools focused and single-purpose

Ideas for Contributions

• Add translation agent support
• Add slide generation (GLM Slide Agent)
• Add streaming support for real-time responses
• Add error retry logic
• Add caching for repeated queries
• Add usage tracking and analytics

📄 License

MIT License - see LICENSE file for details.

🙏 Acknowledgments

• Anthropic for Claude and MCP
• Z.ai for GLM-5 API access
• MCP community for the protocol specification

🔗 Links

• Model Context Protocol
• Z.ai Documentation
• Claude Desktop
• Issue Tracker
• Discussions

💡 Use Cases

Competitive Intelligence

1. web_search("competitor X new features 2024")
2. web_reader(top_results)
3. ask_glm5("Analyze competitor strategy and our response")

Research Synthesis

1. web_search("market trends AI agents", count=50)
2. Multiple web_reader() calls
3. ask_glm5("Synthesize findings into executive summary")

Document Analysis

1. parse_document("https://example.com/contract.pdf")
2. ask_glm5("Extract key terms, risks, and obligations")

Code Generation

1. ask_glm5_pro("Build a React component for user authentication with OAuth")
2. Sonnet integrates into codebase

🆘 Troubleshooting

Tools not appearing in Claude Desktop

1. Check config file path is correct
2. Verify absolute path to index.js
3. Restart Claude Desktop completely (quit, don't just close window)
4. Check Claude Desktop logs for errors

API errors

1. Verify Z.ai API key is valid
2. Check API key has sufficient credits
3. Ensure network connectivity
4. Check Z.ai service status

Empty responses

1. GLM-5 may be rate-limited
2. Try lowering max_tokens
3. Check error logs in terminal
4. Verify prompt is clear and specific

📈 Roadmap

• Add translation agent (40+ languages)
• Add slide generation
• Add image generation
• Add audio transcription
• Add streaming support
• Add response caching
• Add usage analytics
• Add configuration UI
• Add preset prompt templates

Made with ❤️ to help Claude users do more without limits