nvoken
DocumentationIntroduction

Getting started

Introduction

Run durable Agent Turns with optional Conversation continuity and independently selected MemorySpaces.

nvoken is a durable agent runtime for product teams. Your application creates a versioned Agent, then starts a Turn whenever it has work for that Agent. nvoken runs the model-and-tool loop, streams progress, and keeps the accepted work recoverable if a caller disconnects or a worker restarts.

import { Client } from "@deepnoodle/nvoken";
 
const client = new Client();
const agent = await client.agents.create({
  key: "support",
  name: "Support",
  instructions: "Be concise and helpful.",
  model: "anthropic/claude-sonnet-5",
});
 
console.log(await agent.text("Summarize this request.", {
  tenant: "acme",
  user: "alice",
}));

Five independent resources

  • An Agent is stored, reusable behavior in an explicit App, tenant, or user owner namespace. Creating it also creates immutable revision 1.
  • An Agent revision is an immutable behavior snapshot. A Turn records the exact revision it uses; later publication never changes accepted work.
  • A Turn is one durable execution. It always states its tenant and may state an actor user.
  • A Conversation is optional ordered transcript continuity. It does not choose the Agent, actor, or memory.
  • A MemorySpace is optional durable memory selected independently from the Agent and Conversation. It is user-scoped or tenant-namespaced.

A standalone Turn omits a Conversation but keeps the same durable identity, stream, controls, recovery, and result APIs. A conversational Turn selects a Conversation explicitly:

const chat = agent.conversation({
  tenant: "acme",
  user: "alice",
  key: "ticket-483",
  owner: "user",
  memory: { scope: "user", namespace: "support" },
});
 
console.log(await chat.text("What did we decide last time?"));

Your application remains source of truth for users, tenant membership, product data, authorization, billing, and product-level orchestration. nvoken never infers those relationships from resource ownership.

Continue with the quickstart, then read Agents, Turns, Conversations, and MemorySpaces.