CONTROL PLANE FOR MACHINE INTELLIGENCE
One API. Every model.
Control at every
transition.
Modern AI applications are not a single inference request. They are pipelines: a request arrives, a model is chosen, a tool is called, the result is inspected, and the loop continues. Every transition is a decision point — which model handles this, what does this tool call do, is this action safe to allow?
Most applications hard-code a provider and a model, treat tool calling as an implementation detail, and bolt policy on after the fact. When the leaderboard changes, the application must change too. When an agent misbehaves, you find out from the logs.
Sokka makes models interchangeable infrastructure — and makes control a first-class part of the request flow, not an afterthought.
Routing decides which model. The runtime decides how the loop runs. Sokka does both: it picks the right intelligence for each step and applies in-band controls around every agent transition.
Send one request to one API. Sokka classifies the task, estimates complexity, and routes to the best model under your cost, latency, and quality policy.
OpenAI-compatible API
Drop-in replacement for any OpenAI SDK client. Your code does not change when the model does.
Multi-provider routing
Route to any provider from a single endpoint — local, open, or closed models underneath.
Cost-aware routing
Minimize cost while respecting quality constraints. Classifiers inform every decision.
Session stability
Keep the same model for related requests so multi-turn conversations stay coherent.
Fallback chains
Graceful degradation when a model fails or a provider rate-limits.
Usage telemetry
Measure every routing decision, cost, and latency. No blind spots.
An agent is a sequence of decisions: observe, reason, choose tool, authorize, execute, inspect, continue. Sokka puts in-band controls around every transition.
Tool-call inspection
Inspect every tool call before execution. Arguments, targets, and risk are evaluated in-band.
Action authorization
Approve or deny actions in real time. Policy runs inside the loop, not after it.
Agent request routing
Route each agent step to the right model — small for extraction, large for reasoning.
Prompt-injection detection
Detect adversarial inputs in agent loops before they reach a tool.
Audit trails
Complete, replayable record of every agent action and the decision behind it.
Runtime telemetry
Measure everything the agent does — tokens, tools, latency, policy decisions.
Sokka's runtime control pairs with Haiku Guard for classification-based security and FireAgent for isolated execution. Sokka decides what happens; Haiku Guard blocks what shouldn't; FireAgent runs what does.
Sokka and funcroute solve different problems. funcroute routes by content type — it looks at what is attached to a request (image, audio, file) and picks a specialist model. Sokka routes by task — it classifies what the request is asking and applies cost, latency, and quality policy. They compose: run funcroute behind Sokka when you need both content-aware model selection and full routing policy.
Sokka makes routing and control decisions based on the information available at request time. It cannot predict downstream task success with perfect accuracy — benchmark routing policies against your specific workloads before production deployment. Route decisions are only as good as the classifiers that inform them, and policy enforcement depends on correctly configured rules. We recommend defense-in-depth: Sokka routing, Haiku Guard classification, and FireAgent isolation.