Token distillation for coding agents. Same model, leaner payload.
Vachi is an LLM gateway for coding agents. It sits between the agent and the model provider, and applies token distillation to every request before it reaches the model.
Token distillation weighs how much each token contributes to the outcome and rebuilds the payload around what carries real weight. The model receives a leaner, sharper input and does the same work with fewer tokens. Vachi then caches the distilled result so context repeated across many calls isn't paid for twice.
The combined two-step mechanism is called adaptive context caching.
Vachi ships as two products built on that one mechanism. Which one applies depends on how you pay for the model: a subscription plan, or a metered bill.
Go about twice as far on the same Claude plan. Same model, same quality, more work per usage window.
For individuals and small teams on Claude subscription plans (Pro and Max). The constraint here isn't a bill, it's the 5-hour usage window: work stops when the limit is hit. Vachi reshapes each request before it counts against the plan, so the same window carries about twice the work.
npx vachi-claude. One line disables it.Reduce Claude Code costs by 50%.
For companies whose agent traffic is billed per token rather than per seat. Same mechanism, stated as cost: when each request carries about half the tokens, the bill for the same completed work falls by about half.
The enterprise product is bring your own key (BYOK). You add your own Anthropic, OpenAI, or Google API key; Vachi distills each request and calls the provider with that key, so the model bill stays on your own provider account.
Unlike the subscription product, this isn't Claude Code only. Vachi works across coding agents and any client that can point at a custom base URL:
Terms: BYOK, deployable inside your own VPC, zero data retention, hard spend limits, and reversible with one URL change. Pricing is set per engagement — start at vachiai.com/contact.