Model router

The credentials proxy can route by model. Send requests to the virtual host router and each goes to the first rule, in the order you made them, whose models match the request body's model:

harebox proxy rule add --secret-ref anthropic-api-key --host api.anthropic.com --header x-api-key --template '{secret}' --models '["claude-*"]'
harebox proxy rule add --secret-ref openai-api-key --host api.openai.com --models '["gpt-*", "o*"]'
# inside a sandbox:
curl "$HAREBOX_PROXY_URL/router/v1/messages" -H "Authorization: Bearer $HAREBOX_PROXY_TOKEN" -H 'Content-Type: application/json' -d '{"model":"claude-sonnet-5-5", ...}'

A rule can name fallback rules. When its provider answers 429 or 5xx, or can't be reached, the same body goes to each fallback in turn. Use fallbacks for a second key or account with the same provider, or for providers with the same API.

harebox ai usage then reports requests and tokens by model (ai.by_model). Tokens are counted when the provider's response reports them.

The router costs nothing beyond the proxy: $1 for each day it's on, plus $1 per 1,000 requests. Bodies the router reads are limited to 8 MiB.