Run GLM-5.2 in Claude Code with NVIDIA API Key Rotation

Open PowerShell and confirm that Node.js 24 or newer is installed:

node --version
npm --version

The Node.js version must start with v24 or a newer major version. You also need:

• Windows 10 or Windows 11
• An NVIDIA account
• At least one NVIDIA API key
• Internet access to NVIDIA's API catalog

Do not paste an API key into a public command, screenshot, issue, or repository.

Install the current CLI directly from its GitHub repository:

npm install --global github:STEVOHCODER/claude-glm-router

Confirm that the command is available:

claude-glm --version

The package installs Claude Code Router and the compatible Claude Code executable needed by the launcher.

Run the CLI for the first time:

claude-glm

The setup asks you to choose:

1. A provider
2. A model
3. A per-key requests-per-minute limit
4. An optional HTTP/HTTPS proxy
5. One or more API keys

API keys are entered with hidden terminal input. They are not written into the Claude Code Router configuration.

The guided setup creates and manages these local files:

Never upload keys.txt or a provider-key backup. If a key is exposed, revoke it at the provider immediately.

In the provider prompt, choose NVIDIA API Catalog. Select this model when it appears:

z-ai/glm-5.2

You can inspect the current NVIDIA catalog before changing models:

claude-glm models nvidia glm

To select the model later without repeating the complete setup:

claude-glm use-model z-ai/glm-5.2

For NVIDIA, use a conservative local limit of 38 requests per minute for each key. This leaves a small margin below a nominal 40 RPM limit.

When setup asks for a proxy, enter your HTTP/HTTPS proxy URL or leave it blank to use existing environment variables:

$env:HTTP_PROXY = "http://proxy-host:port"
$env:HTTPS_PROXY = "http://proxy-host:port"
$env:NO_PROXY = "localhost,127.0.0.1,::1"

The router keeps local Claude Code traffic off the proxy and uses the proxy only for external provider requests.

Enter your first NVIDIA key during setup. To add another key later, run:

claude-glm add-key

Repeat the command for each key you own. Check the stored keys without revealing their full values:

claude-glm keys

Requests are assigned to available keys in round-robin order. When every key reaches the local rolling limit, the router waits for the next available slot.

Run the built-in checks:

claude-glm doctor
claude-glm status

doctor checks dependencies and configuration. status shows the active provider, model, router health, and per-key rate usage.

If something fails, locate the logs with:

claude-glm logs

A 401 or 403 usually means a key is invalid or not authorized for the model. Run claude-glm setup to replace it.

Start with the recommended NVIDIA profile:

claude-glm mode balanced

Available modes:

Only use resilient when you need fallback behavior because every retry consumes another local rate slot.

Start the background router without opening an interactive Claude session:

claude-glm start

Stop it when needed:

claude-glm stop

To update the CLI, close every running claude-glm session and reinstall:

npm install --global github:STEVOHCODER/claude-glm-router

Start an interactive Claude Code session routed to GLM-5.2:

claude-glm

Run a quick non-interactive test:

claude-glm -p "Reply with exactly: GLM-5.2 is ready"

Monitor the active configuration and local usage:

claude-glm status
claude-glm usage

You can now use Claude Code normally while the local router handles GLM-5.2, API-key selection, proxy access, and rate limits.