GLM Coding Plan glossary

No steps. Split the colliding names, then the cheatsheet rows make sense. Which door to open: learning map.

How the pieces sit

Z.ai / ChatGLM          chat products (#74)
        │
Standard Model API      /api/paas/v4, your own apps
        │
GLM Coding Plan         subscription: GLM only inside listed tools
        ├── Coding agents (Claude Code, Cursor, …)
        ├── General-purpose agents (OpenClaw, …) best-effort
        └── Coding Tool Helper     loader, not an agent

GLM Coding Plan

What it is: “A subscription package designed specifically for AI-powered coding” (Overview).

Why it exists: Use GLM inside Claude Code / Cursor / other listed agents under a credit cap, instead of paying the general API per token.

What it is not: Z.ai chat, ChatGLM, a first-party terminal agent, or a general-purpose API pack.

Supported tools

What it is: The official Coding Agent and general-purpose lists on tool/others. Off-list use does not get plan benefits.

Why it exists: Coding agents are expensive. The vendor keeps plan capacity on named clients plus dedicated endpoints.

What it is not: “Any client that accepts a Base URL.” The China and global lists differ. Do not copy one onto the other.

Coding endpoint

What it is: Plan-only Anthropic / OpenAI URLs on api.z.ai (/api/anthropic, /api/coding/paas/v4, /api/v1).

Why it exists: The same key aimed at /api/paas/v4 is the general API. The Cursor page calls that out explicitly.

What it is not: open.bigmodel.cn. That is the China twin. Keys and hosts stay in their region.

Credits

What it is: A 5-hour cap and a weekly cap. Models burn credits from tokens × multipliers / 10,000. MCP burns credits per call.

Why it exists: One ledger across peak / off-peak, several models, and MCP.

What it is not: Wallet cash. China Overview: when plan credits run out, wait for the next 5-hour window; other packs / balance are not drained.

Best-effort agents

What it is: General-purpose tools (OpenClaw, Hermes Agent, SillyTavern on the global list). Official: served on a best-effort basis; high load may rate-limit. China Overview names this secondary scheduling and gives Coding Agent preemption.

Why it exists: Most plan traffic is coding. Shared capacity prefers coding agents.

What it is not: “OpenClaw is unsupported.” It is listed; it is not first-class.

Team Plan key

What it is: The credential Team members copy from Team Coding Plan > My Plan. “Not interchangeable with other Z.AI's API Keys” (Quick Start).

Why it exists: Seats and org billing must not mix with a personal key.

What it is not: The platform key the admin already had before buying seats.

Coding Tool Helper

What it is: CLI @z_ai/coding-helper (coding-helper / chelper). Loads the plan into Claude Code, OpenCode, Crush, and Factory Droid.

Why it exists: Avoid hand-editing settings.json and MCP.

What it is not: A coding agent. It does not auto-wire Cursor.

Server-side model mapping

What it is: After Claude Code is pointed at the plan, the UI may still show Opus / Sonnet / Haiku while the server maps them to GLM (China Claude page; global page documents the env mapping).

Why it exists: Keep the Claude Code UI. Hard-coded mappings go stale when the plan default model changes.

What it is not: Anthropic inference. /status should show glm-*.

Global site vs China site

What it is: Global = docs.z.ai/devpack + api.z.ai + glm_coding_plan_global. China = docs.bigmodel.cn/cn/coding-plan + open.bigmodel.cn + glm_coding_plan_china.

Why it exists: Different keys, hosts, and allow-lists.

What it is not: Two skins of Z.ai chat.