Fetching from the wire…
Public story · 2026-09-12 · high
litelm keeps 19 provider routes and streaming but drops the proxy server and cost tracking that AI gateway vendors sell as their product.
Why now: The rewrite has 152 points on Hacker News from a September 11 post.
A developer rewrote LiteLLM's provider-routing core in about 2,900 lines with two runtime dependencies, openai and httpx. It went up on Hacker News September 11 under the description "litellm without the bloat."
What got cut matters more than what stayed. litelm drops LiteLLM's Router class, proxy server, response caching, cost tracking, token counting, image generation, and agent framework. Those aren't edge features, they're the core pitch of every AI gateway product on the market: usage visibility and a hosted layer between an app and its model providers.
The rewrite keeps 19 provider routes, message translation, streaming, tool calling, and embeddings. It carries 262 passing tests, 75 of them ported straight from LiteLLM's own suite, ships under MIT, and sits at 162 stars.
Porting 75 of LiteLLM's tests and passing them in a fraction of the code says the routing and translation logic was never the hard part. It shipped bundled with billing and observability tooling that got treated as the same problem, even though nothing in litelm's test suite depends on it.
That split is worth noticing if you're picking an LLM client. Provider routing is small and mechanical enough that a two-dependency fork can replicate it. Cost tracking, caching, and a hosted proxy are the pieces that didn't survive the rewrite, and they're the pieces a gateway subscription is actually selling.
Each link below shares sources, entities, or timing with this story.
Hudson Rock got hold of the archive and counted it. 433,909 files. 118,829 CI runner dumps traced to 2,488 corporate domains. AWS keys, Salesforce client secrets, Slack signing secrets, Azure environment variables, and AI provider API keys belonging to NVIDIA, Volkswagen, Micr...
Y Combinator CEO Garry Tan open-sourced GStack and the repo hit 10,000 GitHub stars in 48 hours. That makes it one of the fastest-growing dev tools of 2026. GStack is a 23-tool MIT-licensed toolkit that turns Claude Code into role-based agents: CEO, Designer, QA, Release Manag...
Anthropic released Claude Fable 5.1 on September 1. Claude Code v2.1.257 made it the default Fable model at 17:53 UTC that day, with a 1M-token context window, $10 per million input tokens, $50 per million output, and $0.25 per million on cache reads (claude-code CHANGELOG). B...
DeepClaude hit 470 points on Hacker News. It swaps Claude Code's API backend to DeepSeek V4 Pro while preserving the full agent loop: file editing, bash execution, git tooling, the whole workflow. DeepSeek V4 Pro scores 96.4% on LiveCodeBench at a fraction of Anthropic's prici...
This Rust harness (+2,585 stars) competes on resource footprint rather than features: 27.8 MB PSS for a single session with local embedding disabled, claimed 13.9× less than Claude Code and 6× less than jcode's own embedding-enabled mode. Time-to-first-frame 14.0ms against a c...
A security scanner. The tool your team trusts to find vulnerabilities. That was the entry point. The TeamPCP campaign compromised Aqua Security's Trivy scanner (a GitHub Action used in CI/CD pipelines), then used that foothold to backdoor LiteLLM's CI/CD pipeline, then pivoted...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.