litelm Reimplements LiteLLM's Call Path in 2,900 Lines and Two Dependencies, Deleting the Proxy, Cache and Cost Tracking

GitHub (metrics verified via the GitHub API; surfaced at 152 points on Hacker News)·medium signalXBlueskyLinkedInCopy link

kennethwolters/litelm took 152 points on Hacker News on September 11 with the description "litellm without the bloat": ~2,900 lines, two runtime dependencies (openai and httpx), 19 provider routes, 262 passing tests including 75 ported from upstream, MIT licensed, 162 stars. It deliberately drops the Router class, proxy server, caching, cost tracking, token counting, image generation and agent framework, keeping only routing, message translation, streaming, tool calling and embeddings. The interesting part for builders is which LiteLLM features turned out to be removable, since cost tracking and the proxy are exactly the layers vendors sell as AI gateways.