claude-additional-models-mcp: Instructions file for Claude Code

CLAUDE.md - GLM-5 MCP Server Project Memory

Project Overview

Purpose: MCP server that integrates Z.ai's GLM-5 (744B parameter model) with Claude Desktop to reduce Claude API consumption by 10x through intelligent task delegation.

Core Value: Solve the "weekly Claude Pro limit exhausted in 2 days" problem by using a hybrid orchestration approach where Claude (Sonnet 4.5 or Opus 4.6) coordinates while GLM-5 handles heavy computational tasks.

Key Metrics:

• 10x reduction in Claude consumption
• 18-30x ROI improvement on Claude Pro subscriptions
• 100% availability (no more multi-day blockages)

Architecture

Hybrid Orchestration Model

User Request
    ↓
Claude / Opus (Parent)
    ├─→ Planning & Coordination (stays in Claude)
    ├─→ File Operations (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)
         ├─→ Complex Analysis (>300 words)
         ├─→ Code Generation (>50 lines)
         ├─→ Research Synthesis
         └─→ Document Processing

Technology Stack

• MCP Server: Node.js + @modelcontextprotocol/sdk
• API Provider: Z.ai (https://api.z.ai/api/paas/v4)
• Models Used:
  • GLM-5 (744B params) - General reasoning & coding
  • GLM-4.6V - Vision & OCR tasks
  • search-prime - Web search engine

GLM-5 Delegation Gate (MANDATORY)

STOP. Run this checklist BEFORE writing code or documentation. This applies to you AND to any Task sub-agents you spawn.

Execution Priority (follow this order)

When facing a task with multiple independent parts:

1. FIRST: Spawn parallel sub-agents — each sub-agent uses GLM-5 for its heavy work. This is the fastest path and MUST be preferred over sequential GLM-5 calls.
2. SECOND: Delegate to GLM-5 directly — when the task is a single unit that cannot be parallelized, use ask_glm5 or ask_glm5_pro.
3. LAST RESORT: Do it yourself — only for orchestration, file I/O, client polish, and responses <100 words. Opus/Claude should almost never generate content directly.