GLM-4.7 Setup

Use GLM-4.7 with agentful for 10x cost savings while maintaining competitive quality.

Why GLM-4.7?

Metric	GLM-4.7	Claude Sonnet 4.5	Difference
Input Cost	$0.60/M	$3.00/M	83% cheaper
Output Cost	$2.20/M	$15.00/M	85% cheaper
SWE-bench	73.8%	77.2%	-4.4%
Math Reasoning	98.6	87.0	+13.3%
Tool Invocation	84.7	Lower	Better
Context Window	200K	200K	Same
Output Capacity	128K	8K	16x larger

Best For:

Cost-sensitive projects
Mathematical/algorithmic tasks
Tool-heavy workflows
Bulk operations
Large output requirements

Quick Start (2 minutes)

Option 1: Using GLM CLI Tool (Easiest)

# Install GLM CLI
npm install -g @xqsit94/glm
 
# Launch Claude Code with GLM
glm --model glm-4.7
 
# That's it! Now run agentful commands:
/agentful-start

Option 2: Environment Variables

# Get API key from https://z.ai
export ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic
export ANTHROPIC_AUTH_TOKEN=your_zai_api_key
 
# Launch Claude Code
claude
 
# Run agentful
/agentful-start

Option 3: Automatic Setup Script

# Download and run setup script
curl -O "https://cdn.bigmodel.cn/install/claude_code_zai_env.sh"
bash ./claude_code_zai_env.sh
 
# Follow prompts to configure
claude

Detailed Setup

Step 1: Get API Key

Visit Z.AI (international) or BigModel (China)
Sign up for free account
Navigate to API Keys section
Generate new API key
Copy the key (starts with zai_... or similar)

Pricing:

Free Tier: Limited requests for testing
GLM Coding Plan Lite: $3/month (~2400 prompts)
GLM Coding Plan Max: ~$10-20/month (production usage)
Pay-as-you-go: $0.60/M input, $2.20/M output

Step 2: Configure Claude Code

Persistent Configuration (Recommended)

Edit ~/.claude/settings.json:

{
  "environmentVariables": {
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "your_zai_api_key",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-4.5-air",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-4.7",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-4.7"
  }
}

Session-Based Configuration

# Set for current session only
export ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic
export ANTHROPIC_AUTH_TOKEN=your_zai_api_key
 
# Launch Claude Code
claude

Per-Project Configuration

Add to your project's .env:

ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic
ANTHROPIC_AUTH_TOKEN=your_zai_api_key

Step 3: Verify Setup

# Start Claude Code
claude
 
# Check status
/status
 
# Should show:
# Model: glm-4.7 (via Z.AI)
# Status: Connected

Test with a simple prompt:

You: What model are you?
Assistant: I am GLM-4.7, running via Z.AI API.

GLM Model Variants

glm-4.7 (Recommended)

Latest and most capable

Released: December 2025
Context: 200K tokens
Output: 128K tokens
Strengths: Math reasoning, tool use, multilingual
SWE-bench: 73.8%
Cost: $0.60/M input, $2.20/M output

# Use in settings.json
"ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-4.7"

glm-4.6

Previous generation

Context: 200K tokens
Strengths: Still very capable, slightly cheaper
SWE-bench: 68.0%
Use Case: Fallback option

glm-4.5-air

Lightweight variant

Context: 128K tokens
Speed: Faster responses
Cost: Lower than 4.7
Use Case: Simple tasks, prototyping

# Good for Haiku-level tasks
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-4.5-air"

glm-4-long

Ultra-long context

Context: Ultra-long (exact limit varies)
Cost: $0.14/M tokens (both input/output)
Use Case: Analyzing massive codebases

Advanced Features

Preserved Thinking Mode

GLM-4.7 can maintain reasoning state across conversation turns:

# Enable via API (future agentful feature)
{
  "clear_thinking": false,  # Preserve reasoning
  "enable_thinking": true    # Enable interleaved thinking
}

Benefits:

42.8% on HLE benchmark (+12.4% improvement)
Consistent architectural decisions
Reduced re-computation
Better for multi-session projects

Large Output Generation

GLM-4.7 can generate 128K tokens in one response (Claude: 8K):

# Use case: Generate entire codebase
You: "Create a full e-commerce application with frontend, backend, and database"
 
GLM-4.7: [Generates 50+ files in single response]
- Complete React frontend (20K tokens)
- Full Express backend (15K tokens)
- Database schemas (5K tokens)
- Docker configuration (3K tokens)
- Documentation (10K tokens)

Context Caching

Reduce costs for repeated context:

# Cached tokens: $0.11/M (vs $0.60/M full price)
# 20-40% cost reduction
 
# Example: Analyzing same codebase multiple times
# First analysis: $0.60/M tokens
# Subsequent: $0.11/M tokens (cached)

Multilingual Support

GLM-4.7 excels at Chinese/English bilingual coding:

# SWE-bench Multilingual: 66.7% (+12.9% over GLM-4.6)
 
You: "创建一个电商网站" (Create an e-commerce site)
GLM-4.7: [Generates code with Chinese comments and English code]

Integration with agentful

All Agents Work Seamlessly

Once GLM is configured, all agentful agents work identically:

# Orchestrator
/agentful-start
 
# Architect
/agentful-product
 
# Backend/Frontend/Tester/Reviewer/Fixer
# All use GLM-4.7 automatically

Recommended Agent Configuration

GLM-4.7 is particularly strong for:

Architect Agent:

Superior at system design
Excellent pattern recognition
Cost-effective for large analysis

Backend Agent:

Strong algorithmic performance
Database optimization
API design

Tester Agent:

Test generation
Edge case identification
128K output = comprehensive test suites

Orchestrator Agent:

Tool orchestration (84.7 τ²-Bench)
Multi-step workflows
Long-running sessions

Cost Optimization Strategies

1. Hybrid Approach

Use GLM for most tasks, Claude for critical code:

# 70% GLM, 30% Claude
# Cost: $1.32/M tokens (56% savings)
 
# 90% GLM, 10% Claude
# Cost: $0.84/M tokens (72% savings)

When to use GLM:

Bulk file analysis
Algorithm implementation
Test generation
Documentation
Refactoring
Tool-heavy workflows

When to use Claude:

Production-critical endpoints
Security-sensitive code
Final code review
User-facing features

2. Model Variant Routing

# Use glm-4.5-air for simple tasks
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-4.5-air"
 
# Use glm-4.7 for complex tasks
export ANTHROPIC_DEFAULT_SONNET_MODEL="glm-4.7"

3. Context Caching

# For repeated codebase analysis
# First run: Full cost
# Subsequent: 81% cheaper (cached context)

Performance Tips

Math-Heavy Tasks

GLM-4.7 outperforms Claude on mathematical reasoning:

# Math benchmark: 98.6 vs Claude's 87.0
 
You: "Implement optimal algorithm for longest increasing subsequence"
 
GLM-4.7:
- Correctly identifies O(n log n) solution
- Provides mathematical proof
- Explains complexity analysis
- Generates test cases
- Optimized implementation

Tool-Heavy Workflows

GLM-4.7 excels at multi-tool orchestration:

# τ²-Bench: 84.7 (beats Claude)
 
You: "Research competitors, analyze market, create presentation"
 
GLM-4.7:
[Tool: web_search] → "AI coding assistants 2025"
[Tool: web_search] → "Market size AI development tools"
[Tool: analyze_data] → Process search results
[Tool: generate_slides] → Create presentation
[Tool: image_search] → Find charts/diagrams
[Success Rate: 85%]

Large Codebase Generation

Use GLM's 128K output capacity:

You: "Generate complete microservices architecture for e-commerce"
 
GLM-4.7: [Generates in ONE response]
- API Gateway (10K tokens)
- Auth Service (15K tokens)
- Product Service (15K tokens)
- Order Service (15K tokens)
- Payment Service (10K tokens)
- Docker Compose (5K tokens)
- Kubernetes configs (10K tokens)
- Documentation (20K tokens)
 
Total: 100K tokens in single response
(Claude would need 12+ separate responses)

Troubleshooting

Connection Issues

# Test API connectivity
curl https://api.z.ai/v1/models \
  -H "Authorization: Bearer $ANTHROPIC_AUTH_TOKEN"
 
# Should return list of available models

Authentication Errors

# Check environment variables
echo $ANTHROPIC_BASE_URL
echo $ANTHROPIC_AUTH_TOKEN
 
# Should output:
# https://api.z.ai/api/anthropic
# zai_...your_key...

Quality Issues

If responses are lower quality than expected:

Check model variant: Ensure using glm-4.7 (not 4.5-air)
Enable thinking mode: Better for complex tasks
Adjust temperature: Lower (0.2) for code, higher (0.7) for creative
Provide more context: GLM benefits from detailed prompts

Rate Limiting

# GLM rate limits (typical):
# Free tier: ~100 requests/day
# Lite plan: ~2400 prompts/5 hours
# Max plan: ~Production usage
 
# If rate limited:
# 1. Upgrade plan
# 2. Use caching to reduce requests
# 3. Add retry logic with backoff

Long Context Degradation

GLM-4.7 occasionally degrades beyond 150K tokens:

# Solution 1: Explicit guidance
You: "Focus on authentication modules in files 50-75"
 
# Solution 2: Chunk large contexts
# Break 200K context into 2x 100K requests
 
# Solution 3: Use glm-4-long variant
export ANTHROPIC_DEFAULT_OPUS_MODEL="glm-4-long"

Switching Back to Claude

Temporary Switch

# Unset GLM configuration
unset ANTHROPIC_BASE_URL
unset ANTHROPIC_AUTH_TOKEN
 
# Set Claude key
export ANTHROPIC_API_KEY=your_claude_key
 
# Launch Claude Code
claude

Permanent Switch

Edit ~/.claude/settings.json:

{
  "environmentVariables": {
    // Remove or comment out GLM config
    // "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_API_KEY": "your_claude_key"
  }
}

Per-Task Switching

# Use GLM for bulk analysis
glm --model glm-4.7
/analyze-codebase
 
# Exit and use Claude for production code
claude
/generate-api-endpoint

Community Tools

GLM CLI Manager

# Install
npm install -g @xqsit94/glm
 
# Features
glm --model glm-4.7          # Launch with specific model
glm token set YOUR_KEY       # Manage API keys
glm token show               # View current key
glm --help                   # See all options

Claude Code GLM Switcher

# Install
npm install -g claude-glm-switcher
 
# Quick launch
claude-glm        # GLM-4.7
claude-glm-air    # GLM-4.5-Air
claude            # Back to Claude

MCP Servers

GLM Math Co-Processor (robertcprice/glm-mcp-server):

{
  "mcpServers": {
    "glm-math": {
      "command": "node",
      "args": ["glm-mcp-server/index.js"],
      "env": { "ZHIPU_API_KEY": "your_key" }
    }
  }
}

Use Claude for orchestration, GLM for math-heavy operations.

Resources

Z.AI Docs: https://docs.z.ai/
GLM-4.7 Overview: https://docs.z.ai/guides/llm/glm-4.7
Function Calling: https://docs.z.ai/guides/capabilities/function-calling
Pricing: https://docs.z.ai/guides/overview/pricing
Hugging Face: https://huggingface.co/zai-org/GLM-4.7
Community Discord: https://discord.gg/agentful

Next Steps

Set up DeepSeek for thinking mode workflows
Configure Gemini for 1M context
Try local models for privacy
Learn cost optimization