Claude Code CLI¶

Claude Code is Anthropic's official agentic coding tool that runs in your terminal. This guide shows how to configure Claude Code to use LLM Gateway.

Prerequisites¶

Claude Code installed (npm install -g @anthropic-ai/claude-code)
LLMGW API token (see User Tokens)

Configuration¶

Claude Code connects to LLMGW via the AWS Bedrock endpoint. Add the following environment variables to your shell configuration file (.bashrc, .zshrc, or similar):

# LLMGW Configuration for Claude Code
export ANTHROPIC_DEFAULT_HAIKU_MODEL="anthropic.claude-haiku-4-5-20251001-v1:0-native"
export ANTHROPIC_DEFAULT_SONNET_MODEL="anthropic.claude-sonnet-4-5-20250929-v1:0-native"
export ANTHROPIC_DEFAULT_OPUS_MODEL="anthropic.claude-opus-4-5-20251101-v1:0"
export ANTHROPIC_BEDROCK_BASE_URL="https://<llmgw-deployment-url>/aws-bedrock"
export ANTHROPIC_AUTH_TOKEN="your-llmgw-token-here"
export CLAUDE_CODE_USE_BEDROCK=1
export CLAUDE_CODE_SKIP_BEDROCK_AUTH=1

After adding these variables, reload your shell:

source ~/.zshrc  # or ~/.bashrc

Environment Variables Explained¶

Variable	Description
`ANTHROPIC_DEFAULT_HAIKU_MODEL`	Model ID for fast/cheap operations
`ANTHROPIC_DEFAULT_SONNET_MODEL`	Model ID for complex tasks
`ANTHROPIC_DEFAULT_OPUS_MODEL`	Model ID for the most capable tasks
`ANTHROPIC_BEDROCK_BASE_URL`	LLMGW AWS Bedrock endpoint
`ANTHROPIC_AUTH_TOKEN`	Your LLMGW API token
`CLAUDE_CODE_USE_BEDROCK`	Enable Bedrock mode
`CLAUDE_CODE_SKIP_BEDROCK_AUTH`	Skip AWS auth (LLMGW handles authentication)

Model Selection¶

Claude Code uses three models:

Haiku model - Used for fast, simple operations (file summaries, quick questions)
Sonnet model - Used for complex coding tasks (code generation, refactoring)
Opus model - Used for the most capable tasks requiring deep reasoning

Cost Saving Tip: For testing purposes, set all models to Haiku to minimize costs. You can upgrade to Sonnet/Opus later when you need better quality for complex tasks.

Available Models¶

Common options:

Model	Use in Variable
Claude Haiku 4.5 (Recommended for testing)	`anthropic.claude-haiku-4-5-20251001-v1:0-native`
Claude Sonnet 4.5 (For production/complex tasks)	`anthropic.claude-sonnet-4-5-20250929-v1:0-native`
Claude Opus 4.5 (Most capable)	`anthropic.claude-opus-4-5-20251101-v1:0`

Verification¶

To verify your configuration is working:

Start Claude Code:
```
claude
```
Ask a simple question to test connectivity:
```
What is 2 + 2?
```
Check the LLMGW Admin Portal to confirm requests are being logged.

Troubleshooting¶

"Invalid API Key" Error¶

If you see authentication errors:

Verify your token hasn't expired
Regenerate a new token if needed from the Admin Portal
Ensure you've copied the full token string
Verify the endpoint URL is correct

"Model not found" Error¶

If you get "model not found" errors:

Check if the selected model ID is correct
Ensure model IDs match exactly (including the -native suffix)

Connection Timeout¶

If you experience connection timeouts:

Check network connectivity to LLMGW
If using a VPN, ensure it allows access to LLMGW endpoints

Slow Responses¶

If responses are slower than expected:

Consider using a faster model for simple tasks
Check if the model is under high load (contact administrator)
Try a different model from the same group

Spend Limit Exceeded¶

If you see an error like:

API Error: 403 Spend limit exceeded for project_user name: <email> · Please run /login

This means you've reached your daily/weekly/monthly spend limit. To resolve:

Wait for the limit to reset (check with your administrator for reset period)
Contact your administrator/PM to increase limits if needed

Best Practices¶

Use environment variables for sensitive configuration like API tokens
Verify connectivity - Check the Admin Portal to confirm requests are being logged
Choose appropriate models - Haiku for simple tasks saves costs
Reload shell after changing environment variables