mcp-glm

An MCP (Model Context Protocol) server that connects Claude Code to GLM/Z.AI models. Designed as a sub-agent that Claude delegates to — minimizing Claude token usage while Claude stays in control as the orchestrator.

Philosophy

Claude is the conductor. GLM is a specialist musician. This MCP lets Claude:

• Delegate entire coding tasks to GLM (fire-and-forget) instead of doing everything itself
• Get second opinions from a different model cheaply
• Offload mechanical work (tests, boilerplate, refactoring) to save Claude tokens
• Review code through a separate model's lens

Features

• Delegate — Fire-and-forget autonomous task execution. Claude sends a task, GLM does all the work, returns a summary. Biggest token saver.
• Review — Structured code review with categorized issues (bugs, security, performance, style)
• Generate — Code generation from specs, matching existing project style
• Chat — Quick questions or second opinions from another LLM
• Vision — Analyze images using GLM-4V-Plus multimodal capabilities
• Embeddings — Generate text embeddings for semantic search and similarity
• Agent (step-by-step) — Fine-grained control when Claude needs to monitor between steps

Prerequisites

• Node.js (v18 or higher)
• npm
• Z.AI API key (get it from z.ai)
• Z.AI Coding Plan (Pro or Max required for GLM-5)

Installation in Claude Code

claude mcp add mcp-glm -e GLM_API_KEY=your-api-key-here -- npx -y github:mathi5/mcp-glm

For global installation (across all projects):

claude mcp add mcp-glm -s user -e GLM_API_KEY=your-api-key-here -- npx -y github:mathi5/mcp-glm

Environment Variables

Variable	Required	Default	Description
GLM_API_KEY	Yes	-	Your Z.AI API key
GLM_MODEL	No	glm-5	Default model for chat/delegate operations
GLM_API_BASE	No	https://api.z.ai/api/coding/paas/v4	API base URL

Available Tools

Primary (recommended)

Tool	Description
glm_delegate	Fire-and-forget — Delegate an entire task to GLM. It reads files, edits code, runs commands autonomously and returns a summary. Best for self-contained tasks.
glm_review	Code review — Send code for structured review. Returns JSON with categorized issues, severity, and suggestions.
glm_generate	Code generation — Generate code from a spec. Pass existing code as context for style matching.
glm_chat	Quick questions — Ask GLM anything. Second opinions, API lookups, brainstorming.

Specialized

Tool	Description
glm_vision	Analyze images with GLM-4V-Plus multimodal model.
glm_embeddings	Generate text embeddings for semantic search/similarity.

Advanced (step-by-step agent)

Tool	Description
glm_agent_start	Start a step-by-step agent session. Use when you need to inspect/intervene between steps.
glm_agent_step	Execute the next step of an agent session.
glm_agent_stop	Stop a session and get the full summary.

Usage Examples

Delegate a task (preferred)

Use glm_delegate with task: "Add unit tests for the calculator module in src/calc.ts"
and context: <paste the file content>

Claude calls once, GLM does everything, Claude reviews the result.

Code review

Use glm_review with the code from src/auth.ts, focus on ["security", "bugs"]

Generate code

Use glm_generate to create a TypeScript interface for the User API response,
with context from the existing types in src/types.ts

Quick question

Use glm_chat to ask: "What's the difference between Promise.all and Promise.allSettled?"

Step-by-step agent (when you need control)

Use glm_agent_start with task: "Refactor the database module"
Then call glm_agent_step and inspect each result
Call glm_agent_stop when done

License

MIT