Key rotation for LLM agents shouldn't require a proxy

I think in-process key management is the right abstraction for multi-key LLM setups. Not LiteLLM, not a Redis queue, not a custom load balancer. Why? Because the failure modes are well-understood. A key gets rate-limited. You wait. You try the next one. Billing errors need a longer cooldown than rate limits. When all keys for a provider are exhausted, you fall back to another provider. This is not a distributed systems problem — it's a state machine that fits in a library. The problem is everyone keeps solving it with infrastructure instead. You spin up a LiteLLM proxy, now you have a Python service to deploy and monitor. You reach for a Redis-backed queue, now you have a database for a problem that doesn't need one. You write a custom rotation script, now it lives in one repo and your three other agent projects don't have it. key-carousel gives each pool a set of API key profiles with exponential-backoff cooldowns. Rate limit errors cool down at 1min → 5min → 25min → 1hr. Billing errors cool down at 5hr → 10hr → 20hr → 24hr. Fallback to OpenAI or Gemini when all Anthropic keys are exhausted. Optional file-based persistence so cooldown state survives restarts. Zero dependencies.