ayounce80/glm-mcp-server

GLM MCP Server

MCP server for Z.ai GLM 4.6 API integration with Claude Code CLI. This allows Claude Code to leverage GLM models for parallel processing, offloading tasks, and cross-validation of outputs.

Overview

The GLM MCP Server enables Claude Code to:

• Offload tasks: Use GLM-4.6 or GLM-4.5-Air for subtasks and agent work
• Parallel processing: Run GLM alongside Sonnet to compare outputs
• Cost optimization: Reduce Sonnet token consumption by delegating to GLM
• Enhanced debugging: Get different perspectives on problems from multiple models

Features

• GLM-4.6 Chat: 357B MoE flagship model for reasoning, coding, and agentic tasks
• GLM-4.5-Air: 106B lighter model for faster responses (when available)
• Reasoning Mode: Extended thinking for complex problems
• Streaming Support: Real-time response generation
• Model Auto-detection: Query available GLM models dynamically

Prerequisites

• Z.ai API Key: Get yours from https://z.ai/manage-apikey/apikey-list
• Python 3.10+: For running the MCP server
• Claude Code CLI: Version 2.0+ with MCP support
• uv: Fast Python package installer (installed during setup)

Setup

1. Install uv (if not already installed)

curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env

2. Clone or Download This Project

cd /home/adam/projects/glm

3. Install Dependencies

uv venv
uv pip install -e ".[dev]"

4. Configure Z.ai API Key

# Option 1: Set environment variable (recommended)
export ZAI_API_KEY="your_api_key_here"

# Option 2: Create .env file
cp .env.example .env
# Edit .env and add your API key

5. Add to Claude Code (User Scope)

claude mcp add --scope user --transport stdio glm-server \
  --env ZAI_API_KEY="${ZAI_API_KEY}" -- \
  /home/adam/projects/glm/.venv/bin/python \
  /home/adam/projects/glm/src/glm_server/server.py

6. Verify Installation

claude mcp list

You should see:

glm-server: /home/adam/projects/glm/.venv/bin/python ... - ✓ Connected

Usage in Claude Code

Once configured, Claude Code will automatically have access to GLM tools. You can also explicitly request GLM usage:

Example Prompts

1. Offload a specific task to GLM:

   Use the glm_chat tool to analyze this code and suggest optimizations
   

2. Compare outputs:

   Check this implementation yourself, and also ask GLM for a second opinion using glm_chat
   

3. Use reasoning mode for complex problems:

   Use glm_reasoning to solve this algorithm problem step-by-step
   

4. Parallel processing:

   Simultaneously (in parallel):
   1. You analyze the bug in module A
   2. Use glm_chat to analyze the bug in module B
   

Available MCP Tools

1. glm_chat

Standard chat completions for general queries, coding, and reasoning.

Parameters:

• prompt (required): The question or task
• model (optional): "glm-4.6" (default) or "glm-4.5-air"
• temperature (optional): 0-1, default 0.7
• max_tokens (optional): Default 4096

Example:

Use glm_chat with prompt "Explain async/await in Python" and model "glm-4.6"

2. glm_reasoning

Extended thinking mode for complex problems requiring step-by-step analysis.

Parameters:

• prompt (required): The complex problem to solve
• model (optional): "glm-4.6" (default) or "glm-4.5-air"
• temperature (optional): 0-1, default 0.7
• max_tokens (optional): Default 8192

Example:

Use glm_reasoning to analyze the time complexity of this recursive algorithm

3. glm_stream

Streaming responses for real-time output (useful for long responses).

Parameters:

• prompt (required): The prompt to send
• model (optional): "glm-4.6" (default) or "glm-4.5-air"
• temperature (optional): 0-1, default 0.7

4. list_glm_models

Query Z.ai API for available models and capabilities.

No parameters required

Available MCP Resources

glm://models

Returns JSON with available GLM models and their specifications.

glm://config

Returns current server configuration and feature status.

Testing

Run Smoke Tests

.venv/bin/python -m pytest tests/test_smoke.py -v

Run Integration Tests (requires API key)

export ZAI_API_KEY="your_api_key"
.venv/bin/python -m pytest tests/test_smoke.py -v

Use Cases

1. Parallel Task Processing

Let Sonnet work on one part of a problem while GLM handles another:

I have two bugs to fix. You fix the authentication issue,
and use glm_chat to investigate the database connection problem.

2. Offloading Agent Tasks

Use GLM-4.5-Air (when available) as a faster alternative to Haiku for subtasks:

Use glm_chat to generate 10 test cases for this function

3. Cross-Validation

Get a second opinion from GLM on critical decisions:

Review my database schema design, then use glm_reasoning
to validate the approach and catch any issues I might have missed

4. Cost Optimization

Delegate simpler tasks to GLM to save Sonnet tokens:

Use glm_chat to write the boilerplate code for these 5 CRUD endpoints

Troubleshooting

Server Not Connected

# Check MCP server status
claude mcp list

# Restart Claude Code or manually restart the server
# The server will auto-restart when you start a new conversation

API Key Not Working

# Verify your API key is set
echo $ZAI_API_KEY

# Test the API key directly
curl -H "Authorization: Bearer $ZAI_API_KEY" \
  https://api.z.ai/api/paas/v4/models

Import Errors

# Reinstall dependencies
cd /home/adam/projects/glm
uv pip install -e ".[dev]"

Architecture

Claude Code CLI
    ↓ (stdio MCP)
GLM MCP Server (Python)
    ↓ (OpenAI-compatible SDK)
Z.ai GLM API
    → GLM-4.6 (357B MoE)
    → GLM-4.5-Air (106B MoE)

Development

Project Structure

glm/
├── src/
│   └── glm_server/
│       ├── __init__.py
│       └── server.py          # Main MCP server implementation
├── tests/
│   └── test_smoke.py          # Smoke tests with mocks
├── pyproject.toml             # Python project configuration
├── README.md                  # This file
├── .env.example               # Environment variable template
└── .gitignore                 # Git ignore patterns

Running the Server Standalone

# The server uses stdio transport, so it's meant to be called by Claude Code
# For testing, you can use the MCP inspector (if installed)
python src/glm_server/server.py

Models

GLM-4.6

• Parameters: 357B MoE (32B active)
• Context: 200K tokens
• Max Output: 128K tokens
• Best for: Reasoning, coding, agentic tasks, tool use

GLM-4.5-Air

• Parameters: 106B MoE (12B active)
• Context: 128K tokens
• Best for: Faster responses, general chat, lighter tasks

Contributing

1. Run tests before committing:

   .venv/bin/python -m pytest tests/ -v
   

2. Follow the existing code style

3. Update tests for new features

License

MIT

Resources

• Z.ai Platform: https://z.ai
• Z.ai API Docs: https://docs.z.ai
• MCP Documentation: https://docs.claude.com/en/docs/claude-code/mcp
• Claude Code: https://code.claude.com