Kultivait.ai is a local-first routing proxy. Point your tools at one OpenAI- or Anthropic-compatible endpoint; every prompt is weighed by a local embedding model and routed to the cheapest model that can carry it — your own machine first, the cloud only when it earns its cost.
Most setups point every prompt at one premium cloud model, chosen by habit rather than fit. But the bulk of day-to-day AI work — renames, summaries, drafts, commit messages, classification — doesn't need a frontier model. You're paying frontier prices anyway.
Per-token pricing means costs scale with usage, and agentic tools multiply usage. What starts as a rounding error becomes a real line item.
A human picked one model for everything. Simple prompts burn premium tokens; the choice never adapts to the prompt in front of it.
Sending trivial prompts to massive data-center models burns energy that a small local model would sip. Overkill has a footprint.
Point your tools at Kultivait.ai instead of a single provider. The router reads each prompt and sends it where it's best served — local and free by default, premium cloud when the problem earns it.
Swap one endpoint, classify in milliseconds, and tally every routing decision in a local savings ledger.
One-command install, OpenAI + Anthropic compatible endpoints, CLI reference, and agent integration.
Less compute, less carbon — why local-first routing is the sustainable choice.
Type a task you'd hand to AI — or pick an example — and see where Kultivait.ai would send it, and what each path costs.
Multiply that by every prompt your agent sends in a session — see the savings math.
This demo is a simplified sketch. The real router classifies with a local embedding model (nomic-embed-text) against seed-prompt centroids — no cloud call decides whether to make a cloud call.
Every routing and distillation decision in Kultivait.ai was chosen by an experiment you can rerun, and every claim below links to how it was measured.
On held-out prompts, the embedding classifier routed 24 of 24 correctly with zero dangerous misroutes — no cloud-worthy work sent to a weaker model.
The context distiller keeps every planted fact while cutting transcripts to 61% of their original tokens — the winning model was chosen from a 5-model eval.
OpenAI-compatible and Anthropic-compatible endpoints, both streaming, with client-side tool calls passed through. Your tools don't change; your bill does.
Details, tables, and the full command reference live on the developers page; the code lives on GitHub.
Kultivait.ai is free during early access. If you run local models with ollama, you're two minutes from routing.
curl -fsSL https://kultivait.ai/install.sh | sh
Not ready to install? Leave an email and we'll send occasional progress notes — no spam, just harvests.
Running AI across a team? Join the Kultivait.ai for Teams waitlist.