Show HN: Butter – A Behavior Cache for LLMs (www.butter.dev)

Butter is a behavior cache for LLMs that detects recurring patterns in model responses and serves those outputs directly so you don’t need to pay for the same completions repeatedly. It functions as a drop‑in Chat Completions API endpoint (you simply repoint your client to Butter’s base URL), so it works with popular tooling like LangChain, Pydantic AI, LiteLLM, Helicone and many agent frameworks. Butter is deterministic, meaning cached responses are replayed consistently — useful for agents and back‑office automation where reproducible behavior matters — and the service is live with a demo available. For AI/ML teams and autonomous agents that perform repetitive tasks (data entry, research queries, scripted tool use), Butter can cut token costs and reduce latency by returning cached outputs instead of hitting the model every time. Key technical implications include lower API spend, faster response times, and reproducible agent behavior; tradeoffs to consider are cache coverage, invalidation freshness, and ensuring cached responses remain correct as contexts or downstream tools change. The company’s pricing is aligned to savings (5% of the tokens saved) and the product is currently free to try, making it an attractive cost-optimization layer for production LLM deployments.