Glm Mcp

What is Glm Mcp?

Glm Mcp is an MCP server that wraps GLM (Zhipu / Z.ai) as a ~10x cheaper delegate for AI coding agents. It exposes an Anthropic-compatible /v1/messages endpoint and provides four tools. It is designed for users who want an expensive main model (e.g., Claude Opus) to orchestrate and review work while GLM does the actual coding and text generation.

How to use Glm Mcp?

Install via one of three packages: glm-mcp-claude for Claude Code, glm-mcp-copilot for GitHub Copilot/VS Code, or the standalone glm-mcp for any MCP client (Cursor, Windsurf, Glama, etc.). After installing, restart the client and use the tools glm_agent, glm_delegate, glm_recommend, and glm_status. A GLM API key (from Z.ai) is required for actual GLM calls.

Key features of Glm Mcp

• Delegates coding tasks to GLM at ~10x cost savings.
• Four tools: agent, text delegate, recommendation, and status.
• Peak-aware model routing with configurable cost bias.
• Git checkpoint revert and dry_run for safe previews.
• Live progress notifications and a usage ledger (usage.jsonl).
• Works with Claude Code, GitHub Copilot, and any stdio MCP client.

Use cases of Glm Mcp

• Use GLM as a coding agent to read, write, edit files and run bash commands.
• Generate text with glm_delegate when no file access is needed.
• Preview a full diff before applying changes with dry_run: true.
• Track GLM token usage and estimated costs via the ledger and glm_status.
• Route simpler tasks to GLM while reserving the main model for orchestration and review.

FAQ from Glm Mcp

What is the cost advantage of using Glm Mcp?

GLM tokens are approximately 10x cheaper than the main model (e.g., Claude Opus). The main model pays only for orchestration and review, while the actual coding work is billed on cheap GLM tokens.

How can I verify that GLM is really being used?

Every GLM call is logged to a usage.jsonl ledger on disk. The glm_status tool prints cumulative ledger totals (calls, tokens, per-model counts), and each glm_agent run outputs a GLM STATS block showing the model used and token counts.

What happens during China peak hours?

Peak hours are 14:00–18:00 (UTC+8). The router adjusts model selection to avoid surcharges on glm-5.x models. If a no-surcharge model like glm-4.7 is listed in `GLM_PEA