@nexural/sdk

Shared helpers for warehouse MCP servers, recipes, and forged apps.

llmClient — the cost wrapper every emitted app uses

Per ADRs 0007 + 0010, recipes that import provider SDKs directly fail federation-conformance. Use llmClient() instead.

import { llmClient } from "@nexural/sdk";

const client = llmClient({
  appName: "my-rag-app",
  recipeName: "saas-rag-chat",
  costEnvelope: {
    /* from recipe.yaml */
  },
  modelChain: ["anthropic:opus", "openai:flagship", "ollama:llama-large"],
  usageTracker, // your Postgres / SQLite-backed counter
  telemetry, // emits cost_event to nexural-meta
  providers, // Map of provider id → caller
});

const result = await client.invoke({
  userId: "user-123",
  messages: [{ role: "user", content: "..." }],
  inputTokenEstimate: 1500,
  maxOutputTokens: 500,
  streaming: false,
});

Hard caps from costEnvelope.hard_caps:

• per_request_usd — pre-flight reject; mid-stream abort
• per_user_per_day_usd — block individual users
• per_app_per_day_usd — circuit-break entire app

All breaches emit cost_event telemetry. Users hashed (sha256) — raw IDs never stored.

checkDecay + sha256Hex

import { checkDecay, sha256Hex } from "@nexural/sdk";

const status = checkDecay("2026-01-01", 90);
// → { status: "stale" | "fresh" | "quarantined" | "auto-deprecate", ... }

const argsHash = sha256Hex(JSON.stringify(args)); // for telemetry