Coding Tools (GLM Coding Plan)

go-z-ai coding configures third-party coding assistants to use your GLM Coding Plan instead of their default provider. It's a Go port of Z.AI's official @z_ai/coding-helper ("chelper") CLI, sharing the same credential file so the two tools can be used interchangeably.

Supported tools

Run go-z-ai coding tools to see install status and exact resolved paths on your machine.

Plans

Pick whichever matches where your GLM Coding Plan subscription lives.

Quickstart

# 1. Store and validate your GLM Coding Plan key (one-time)
go-z-ai coding auth glm_coding_plan_global YOUR_KEY

# 2. Load it into a tool
go-z-ai coding load claude-code
# tool IDs: claude-code, opencode, crush, factory-droid, cursor
# aliases also work: claude, droid, factory

# 3. Check everything's wired up
go-z-ai coding status
go-z-ai coding doctor

Credentials live at ~/.chelper/config.yaml (byte-compatible with the official Node helper) — coding auth writes there once, and coding load reads from it for every tool unless you pass --key/--plan overrides.

To stop using Z.AI for a tool without losing your stored credential:

go-z-ai coding unload claude-code

This removes only the Z.AI-specific fields it added; it does not touch the rest of your existing config file.

Claude Code: model mapping

The official helper only sets ANTHROPIC_AUTH_TOKEN, ANTHROPIC_BASE_URL, API_TIMEOUT_MS, and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC. This client goes further by default and also maps Claude Code's model tiers to specific GLM models via ANTHROPIC_DEFAULT_*_MODEL, matching Z.AI's documented recommendation:

Override any tier, or opt out entirely:

go-z-ai coding auth glm_coding_plan_global YOUR_KEY \
  --sonnet glm-5.2 --opus glm-5.2

go-z-ai coding auth glm_coding_plan_global YOUR_KEY --no-model-mapping

Also configurable, all optional (0/omitted = don't set the env var):

These flags are persistent on the coding command, so they apply the same way to auth, load, and reload.

Key management

go-z-ai coding auth revoke              # clear the stored key, keep the plan choice
go-z-ai coding auth reload <tool>       # re-push stored creds into a tool
go-z-ai coding load <tool> --key OTHER_KEY --plan glm_coding_plan_china  # one-off override

By default, coding auth validates a new key against the API before storing it (a real /models call). Skip that with --no-validate if you want to store a key offline (e.g. scripting a machine you haven't network-tested yet).

Vision MCP server

The official @z_ai/coding-helper wizard has a "manage MCP services" step that this client didn't replicate until now: Z.AI ships its own Vision MCP Server (@z_ai/mcp-server) — screenshot OCR, error-screenshot diagnosis, diagram/chart understanding, and general image/video analysis via GLM-4.6V, launched on demand via npx. coding mcp registers it in whichever tool you're using:

go-z-ai coding mcp add claude-code     # uses the stored API key
go-z-ai coding mcp add crush --key OTHER_KEY
go-z-ai coding mcp status              # which tools have it configured
go-z-ai coding mcp remove claude-code

Requires Node.js. The server itself runs via npx -y @z_ai/mcp-server — Z.AI's own docs currently recommend Node.js 22+, though the npm package only declares an 18+ requirement. coding mcp add/doctor warn (don't block) if npx isn't found on PATH, since the config is valid the moment Node.js becomes available.

The MCP config file often isn't the same file as your GLM credential. Two of the five tools keep MCP servers in a separate file from provider/API settings:

Cursor isn't explicitly listed as a supported Vision MCP client in Z.AI's own docs — this uses the generic MCP shape Cursor documents for any server, which should work but hasn't been Z.AI-confirmed for this specific server.

Doctor

go-z-ai coding doctor

Checks: is a credential stored, does it look well-formed, which supported tools are installed on PATH, and which of those already have a Z.AI configuration (including the Vision MCP server, if registered). Good first step when something isn't working.

Compliance & usage policy ⚠️

Z.AI's coding endpoint (/api/coding/paas/v4) is restricted by the usage policy to "officially supported tools," and "SDK-based access" is explicitly prohibited. Three violations result in an account ban. Unidentified third-party clients are indistinguishable from prohibited access at the server (pi#4187).

go-z-ai mitigates this in two ways:

1. The coding subcommand wires officially-supported tools (Claude Code, OpenCode, Crush, Factory Droid, Cursor) into their native config formats — exactly what Z.AI's own @z_ai/coding-helper does. The wiring itself is the supported path; go-z-ai just automates it from a Go binary.
2. Every request carries an identifying User-Agent: go-z-ai/<version> header (overridable via Config.UserAgent for downstream apps, proxies, and MCP servers that need their own identifier). This is the minimum hygiene that distinguishes go-z-ai from anonymous/prohibited access.

What this does not do:

• It does not make go-z-ai an "officially supported tool" — only Z.AI can confer that status. Until then, using the coding subcommand to wire a supported tool is the compliant path; using pkg/client directly against /api/coding/paas/v4 from a custom integration is at the user's own risk.
• It does not spoof or evade detection. The header honestly identifies the client; the goal is to be a good citizen, not to hide.

If you are building a downstream tool, proxy, or MCP server on top of pkg/client, set a distinct Config.UserAgent (e.g. "my-tool/1.0 (go-z-ai)") so Z.AI can identify your traffic separately and so you inherit go-z-ai's good-citizen default rather than weakening it.