Usage with Claude Code

This guide shows how to run Claude Code through ekai-gateway for multi‑provider routing, unified configuration, and detailed usage analytics.

What You Get

• Single endpoint for Anthropic, OpenAI, xAI, and OpenRouter

• Automatic model routing based on model name

• Usage and cost analytics at http://localhost:3000

• Works with standard Claude Code env vars

Prerequisites

• Node.js 18+ and npm

• API keys for any providers you plan to use

Install and Run the Gateway

git clone https://github.com/ekailabs/ekai-gateway.git
cd ekai-gateway
npm install
npm run dev

Configure Environment

Copy and edit .env with your keys:

cp .env.example .env
# Edit .env and set any of:
# ANTHROPIC_API_KEY=...
# OPENAI_API_KEY=...
# XAI_API_KEY=...
# OPENROUTER_API_KEY=...
# Optional
# PORT=3001        # Gateway API port
# UI_PORT=3000     # Dashboard port

Point Claude Code at the Gateway

Claude Code uses ANTHROPIC_* variables. Set the base URL to the gateway and choose a model name; the gateway will route it to the right provider.

# Point Claude Code to ekai-gateway
export ANTHROPIC_BASE_URL="http://localhost:3001"

# Pick a model (examples)
export ANTHROPIC_MODEL="claude-sonnet-4-20250514"   # Anthropic
export ANTHROPIC_MODEL="gpt-4o"                     # OpenAI
export ANTHROPIC_MODEL="grok-code-fast-1"           # xAI
export ANTHROPIC_MODEL="moonshotai/kimi-k2"         # OpenRouter

# Start Claude Code
claude

Notes

• Set only one ANTHROPIC_MODEL at a time per session.

• You can run multiple Claude Code sessions with different models by setting different envs per terminal.

Verify and Monitor

• Hit the dashboard at http://localhost:3000 to see token usage, costs, and model/provider breakdowns.

• Check gateway logs in the terminal for request details.

Troubleshooting

• 401/403 errors: verify the relevant API key is present in .env and the provider allows the chosen model.

• Connection refused: ensure npm run dev is active and ANTHROPIC_BASE_URL matches the gateway port.

• Model not found: confirm the model name is supported by its provider and spelled correctly.