# Introduction (/) Zumik sits between your agent and the model providers. It measures how much of your inference work repeats, keeps the stable parts of a request as reusable state, and routes each generation to the cheapest reliable path across managed providers, your own keys, or your own cloud. Agent applications resend the same system instructions, tool definitions, schemas, policies, and conversation history on nearly every call. A large fraction of the input stays fixed while a small suffix changes. That repeated work costs latency and money. Zumik makes it visible, makes the reusable parts addressable, and reports how much reuse you actually captured, not just how much was possible. ## The control plane thesis [#the-control-plane-thesis] Zumik is not a model marketplace and not a new inference engine. It is a control plane: one internal execution system behind two public API surfaces, with explicit state, reproducible routing, and honest measurement. A workload diagnostic scores reuse opportunity from metadata traces and recommends the lowest-complexity profile that improves cost or latency. A reusable handle is not a cache hit. Zumik reports both, with an evidence level on every realized number. Model aliases resolve to immutable releases. Snapshots pin state. Every response records exactly what it ran against. Start on managed providers. Bring your own keys or your own cloud only where replay data proves the move is worth it. ## Who it is for [#who-it-is-for] Zumik fits teams whose traffic has real locality: coding agents that resend repository instructions and tool registries, research agents with long task histories, enterprise copilots with tenant instructions and domain documents, support automation, and RAG platforms that reuse documents and schemas. It is a poor fit for one-off short prompts, consumer chat with weak locality, pure API aggregation, or teams shopping only for the lowest nominal token price. If your workload does not repeat, there is nothing for Zumik to optimize, and the diagnostic will say so. ## Two API surfaces [#two-api-surfaces] Zumik exposes two public APIs over one execution system. Exact OpenAI request and response shapes. Swap the base URL and your existing client works unchanged. Proprietary behavior rides on optional `Agent-*` headers, never in the JSON body. Artifacts, bundles, sessions, branches, snapshots, diagnostics, replay, and signed purge receipts as first-class objects. The two surfaces interoperate. You can migrate an OpenAI client to `/v1` today with no body changes, then adopt `/v2` state incrementally where it pays off. The compatibility path stays correct whether or not you use the native one. ## Where to go next [#where-to-go-next] First request in five minutes: swap the base URL, create an artifact, run a diagnostic. API keys, bearer auth, per-key budgets, and console sessions. The object model, the three identity layers, aliases, and purge. Every term Zumik uses, with a one-line definition and a link to the detail. Prepaid credits (from $5), budgets, and the reuse credit. Tenant isolation, data privacy, retention, and the vulnerability checklist.