glm-mcp — run GLM as a cheap delegate for your AI coding agent
Use the GLM model (Zhipu / Z.ai) as a ~10× cheaper delegate inside your AI coding tool.Two editions, the same GLM MCP server underneath — pick your editor:
Pick your editor
🟠 Claude Code → claude/ |
⚫ GitHub Copilot (VS Code) → copilot/ |
|
|---|---|---|
| npm package | glm-mcp-claude |
glm-mcp-copilot |
| Install | npx glm-mcp-claude --key YOUR_ZAI_KEY |
npx glm-mcp-copilot --key YOUR_ZAI_KEY |
| Integration | MCP server + full-tool glm subagent + auto-delegation hook + glm-code full-GLM launcher |
MCP server in agent mode + .github/copilot-instructions.md |
| Tools | glm_agent · glm_delegate · glm_recommend · glm_status |
same |
| Docs | claude/README.md | copilot/README.md |
Copilot has no subagents/hooks, so that edition uses MCP tools + an instructions file instead of asubagent + hook. Everything downstream (GLM agent loop, peak-aware routing, cost bias, token cap,usage ledger, dry-run oversight) is the same server, so behavior is identical once a tool runs.
Global vs per-project
- Claude Code installs globally by default (user-scoped MCP server +
~/.claudeagent/hook) — works in every project. - GitHub Copilot — add
--globalto configure every workspace at once (writes VS Code's usermcp.json+settings.json); omit it to set up just the current project:npx glm-mcp-copilot --global --key YOUR_ZAI_KEY
What it does (both editions)
- Offloads well-specified coding work to GLM —
glm_agentlets GLM read/write/edit/run your repodirectly, on GLM tokens (~10× cheaper). Your main model orchestrates + verifies. - Peak-aware model picks, a cost bias that keeps GLM the default, a token cap toggle, ausage ledger (proof of GLM tokens), and dry-run oversight (preview a diff before applying).
Requirements
- A Z.ai / Zhipu GLM Coding Plan API key — https://z.ai (the only paid key needed).
- Node.js ≥ 18, plus the editor (Claude Code app, or VS Code + Copilot with agent mode).
License
MIT © djerok · Canonical repo: https://github.com/djerok/glm-mcp
