Claude Code on your own GPU

What you need

• A WAI gateway key from Dashboard -> Settings -> API Keys.
• A node online if you want local serving. For coding, deploy a coder model such as Qwen Coder from the Models page.
• A context window large enough for agent work. Start at 32k when your hardware can fit it.
• A Claude-shaped model alias, usually claude-local-code, created in Dashboard -> Models -> Model Mapping tab.

Quickstart

Claude Code uses Anthropic environment variables, not OpenAI ones. Set the base URL to WAI's /anthropic surface and use your WAI gateway key as the Anthropic auth token.

# 1. Use your WideAreaAI gateway key from Dashboard -> Settings -> API Keys.
export ANTHROPIC_BASE_URL="https://wideareaai.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="wai_sk_..."

# 2. Let Claude Code list Claude-shaped models from the gateway.
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1

# 3. Use a Claude-shaped alias that you create in Dashboard -> Models -> Model Mapping tab.
claude --model claude-local-code

$env:ANTHROPIC_BASE_URL = "https://wideareaai.com/anthropic"
$env:ANTHROPIC_AUTH_TOKEN = "wai_sk_..."
$env:CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY = "1"

claude --model claude-local-code

Create the model alias

Claude Code expects Claude-style model names. In WAI, create an alias so Claude Code asks for a Claude-shaped id while the gateway can serve a local model first.

With this setup, Claude Code can keep using claude-local-codewhile WAI decides whether the actual request lands on your node, a cloud Anthropic model, or another configured fallback.

Verify the gateway

Before opening a real repository in Claude Code, verify the Anthropic endpoint with a tiny request.

curl https://wideareaai.com/anthropic/v1/messages \
  -H "Authorization: Bearer wai_sk_..." \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-local-code",
    "max_tokens": 128,
    "messages": [
      {"role": "user", "content": "Reply with one sentence from the WAI Anthropic gateway."}
    ]
  }'

curl -N https://wideareaai.com/anthropic/v1/messages \
  -H "Authorization: Bearer wai_sk_..." \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-local-code",
    "max_tokens": 128,
    "stream": true,
    "messages": [{"role": "user", "content": "Stream three short bullets."}]
  }'

curl https://wideareaai.com/anthropic/v1/models \
  -H "Authorization: Bearer wai_sk_..."

Supported Anthropic endpoints

Tools and routing behavior

Claude Code sends tool definitions and tool results through the Anthropic Messages API. WAI translates those into OpenAI-compatible tool calls before routing to your node or cloud fallback. Local model quality depends heavily on whether the model supports function/tool calling and whether the node was started with a large enough context window.

You can still use WAI routing controls. Add X-WAI-Node to target a specific node when testing, or let the gateway rank ready nodes and fail over according to your account settings. Response headers such as X-WAI-Served-By, X-WAI-Served-Model, andX-WAI-Alias-Id show where a request actually ran.

Token counting

Claude Code may call /v1/messages/count_tokens. WAI answers that endpoint so the client does not fail, but the count is an estimate. Actual request logging and credit billing still come from the model response usage when the upstream reports it.

curl https://wideareaai.com/anthropic/v1/messages/count_tokens \
  -H "Authorization: Bearer wai_sk_..." \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-local-code",
    "messages": [{"role": "user", "content": "Count this."}]
  }'

Troubleshooting