@adaptivemcp/routing

Model selection and cost optimization for Adaptive MCP.

Status: published to npm as @adaptivemcp/routing. It is the client-side executor of policy the server governs.

routing reads learned stats from the store (@adaptivemcp/memory) and emits model and routing recommendations. For example, steer a flaky or expensive tool toward a cheaper/faster model once the evidence supports it.

Usage

import { MemoryStore } from "@adaptivemcp/memory";
import { Router } from "@adaptivemcp/routing";

const memory = new MemoryStore();
const router = new Router({
  memory,
  models: [
    { id: "gpt-5-mini", costWeight: 1, latencyWeight: 1 },
    { id: "gpt-5", costWeight: 4, latencyWeight: 0.6 },
  ],
  budget: { perToolLimit: 0.05 },
});

router.routeAll(); // writes `model` + `routing` recommendations into the store

What it does

• routeAll() / routeTool(toolName): derive routing recommendations from observed stats.
• Emits model and routing recommendations into the store.
• Warns when projected cost crosses ~80% of a configured budget.

Defaults

| Option | Default | | --- | --- | | minInvocations | 10 |

License

Released under the MIT License. Copyright (c) 2026 Kemal Elmizan.