Solo Developer

PennyRouter — Every token at its best price. PennyRouter is a blazing-fast, resource-frugal AI gateway written from scratch in Rust. It speaks the OpenAI API on the way in (change only base_url), routes each request to the best-matching provider with one local lookup, and streams the answer back — on roughly 1 CPU / 1 GB RAM, up to 1M-token requests. The edge is the continuously re-ranked provider index behind the lookup: price (input/output/cache blended), latency, throughput, availability, precision class (FP8 != BF16), promos, and offline-judged quality — not a static sort order. Sessions pin to one provider so prompt cache survives. Every response carries x-pennyrouter-provider, so the choice is verifiable. What you type: virtual tiers openrouter/pair/free | /low | /think | /lowest, decision tiers, model pins (model/<id>), fallback chains (free/fallback/<tier>), suffix tags (/low_cost, /speed, /privacy), provider pins (modal/<model>, baseten/<model>), hard budgets (provider.max_price), tool-aware routing, and an opt-in smart classifier (smart[/max]). Resilience is honest: one careful retry, header-driven 429 cooldowns, clean 503 load-shed, OpenAI-shaped errors, fail-closed everywhere. BYOK throughout — the gateway stores no serving keys. Docker local, Terraform/Hetzner cloud. MIT. Built during the hackathon (Oct 26–Nov 4): Iters 19–29 — lowest tag, gzipped error decode, free-fallback chains + 429 headers, tool-aware routing, standby recovery, decision tier, OpenRouter-native interface, smart classifier, decisions route, pinned-model routes, routing inspector — plus the demo video. Measured overhead +0.26 ms p50 (JSON) / +0.16 ms (SSE) vs direct-to-mock.