Free where it should be free
Everything an individual or small team needs costs nothing — full engine, real production use. We'd rather you grow into a conversation than resent a paywall.
About Crowkis
The cache that knows when to say no.
Crowkis is a Redis-compatible cache, written in Rust, that understands what your LLM is being asked — and only reuses an answer when it can prove the reuse is safe. It exists because the cheapest token is the one you never spend twice, and the most dangerous cache is the one that can't tell two questions apart.
Why it exists.
Every LLM team we met was paying twice for the same answers — the same questions, rephrased all day, billed at full price every time — and hoping a vector database would save them. It doesn't. Match by raw similarity and you serve “cancel my subscription” the answer meant for “pause my subscription.” Match by exact bytes and you never hit at all.
Crowkis is the cache built for that gap: it reuses aggressively when meaning and structure agree, and refuses when they don't. We wrote the storage engine from scratch in Rust, gave it an immune system against poisoned writes, and shipped it as a single signed image that never phones home — because the component holding every question your customers ask should be the most trustworthy thing in your stack, not the least.
The bill gets lighter. The answers get faster. And the cache earns its place by being the part of the stack you stop worrying about.
~0
lines of Rust
0
integration tests
0
intelligence systems
0
protocol surfaces
What we believe
Six principles, no compromises.
Safety is the product
A cache that reuses LLM answers is a trust system. Saying 'no' to an unsafe hit is the feature — five checks gate every read, five stages gate every write.
No supply chain to attack
One signed Rust binary. No Python, no PyPI, no dependency tree in the runtime image. The attack class that hit the ecosystem in 2026 can't apply by construction.
Offline, always
Nothing phones home. The license check is local Ed25519. Your customers' questions never leave your network — air-gapped deployments are first-class.
Built in Rust, on purpose
A custom LSM engine, an in-process vector index, no garbage collector in the read path. Sub-millisecond hits aren't a benchmark trick; they're the architecture.
Boring on purpose
The highest compliment infrastructure can earn is 'we never think about it.' One image, stable on-disk format, fail-open everything. Upgrades are a non-event.
Contact
Building with LLMs? Let's talk caches.
Evaluating Crowkis, stuck on a deploy, weighing Enterprise, or just want to compare notes on semantic caching — send a message and a human will get back to you.
Send us a message
Fill this in and it'll open your mail app, ready to send to us.