Hey PH 👋 I'm Kushal, maker of MorrowCache. The pain: same user question, different wording—or an agent that rephrases until the answer "looks done"—and you pay for every chat completion again. Vector-ish caches miss paraphrases or merge the wrong prompts. MorrowCache is a local OpenAI-compatible proxy. Point your client at http://127.0.0.1:8080/v1, keep your upstream model. Before calling the model, a same-intent adjudicator (TypeSafe Jev by default, or local Kev/Laya) decides whether to reuse the cached answer. Not "similar text" as the only signal. What I'd call out upfront: • Fail-open — if the judge errors, the real model still runs • Freshness — time-sensitive asks can refuse a stale HIT • Streaming + tool calls bypass the cache (by design in v0) Quick try (~5 min, Node 22.5+): npx @kushalicious/jevcache@latest start Demo with no real API key: MOCK_UPSTREAM=1 npx @kushalicious/jevcache@latest start --demo Then send a prompt + a paraphrase → expect MISS then HIT (X-Jevcache headers or /stats). Easiest in Cursor / Claude Code: copy the setup prompt from https://morrowcache.vercel.app/a... Recorded pair on the site: 2817ms MISS → 423ms HIT at intent 0.94 — one example, not a guarantee for every workload. I'd love feedback from anyone running agents or support bots: where would same-intent caching help, and where would you never trust it? Happy to dig into architecture in the comments.

MorrowCache

Skip the chat call when the question is the same

Skip the chat call when the question is the same