Nexus

Enterprise-ready agent framework for TypeScript. Describe agents in config. Wire who is calling and where data lives at run time. Call run().

No global LLM settings. No shared agent singleton. Built for multi-tenant SaaS apps.

Beta — under active development. Import as nexus-enterprise-agent.

Why Nexus?

• SaaS-native — tenant, user, API keys, and storage passed per request
• Context-first (RCS) — compresses old tool results inline; no extra LLM calls
• Config-driven — YAML manifests for teams; TypeScript for wiring and tools
• Provider-agnostic — bring your own model; pluggable storage
• Voice (client-side) — cascaded STT → LLM → TTS with mock adapters; GPU media servers stay in the Python package

Install

npm install nexus-enterprise-agent
# or from GitHub (builds via prepare)
npm install github:gowrav-vishwakarma/nexus-enterprise-agent-js

Optional peers (install what you use):

| Package | Adds | |---------|------| | @ai-sdk/openai, @ai-sdk/anthropic, … | LLM clients via the Vercel AI SDK | | @libsql/client, postgres, ioredis | Storage adapters (when implemented) |

Contributors:

git clone https://github.com/gowrav-vishwakarma/nexus-enterprise-agent-js.git
cd nexus-enterprise-agent-js
npm install
cp .env.example .env   # set your LLM API key
npm test

Text agents (YAML)

Three files: manifest, prompts module, short runner script.

npx tsx examples/orchestration/run_team.ts "Analyze Q4 revenue"

• Manifest: examples/orchestration/research_team.yaml
• Prompts: examples/orchestration/research_team_prompts.ts

TypeScript API (same agents, built in code):

import { AgentRunner, createRunContext, parseAgentConfig } from "nexus-enterprise-agent";
import { ToolRegistry } from "nexus-enterprise-agent";

const runner = new AgentRunner({
  config: parseAgentConfig({
    name: "researcher",
    llm: { provider: "openai", model: "gpt-4o", api_key: process.env.OPENAI_API_KEY },
    persona: { role: "Researcher", goal: "Answer questions" },
  }),
  toolRegistry: new ToolRegistry(),
  runContext: createRunContext({ tenantId: "acme", userId: "u1", sessionId: "chat-1" }),
});
const result = await runner.run("What is AI?");

Walkthrough: docs/getting-started.md (YAML) · docs/getting-started-typescript.md (TypeScript)

Voice agents (client-side)

Cascaded voice: VAD → STT → LLM → TTS. GPU servers stay in Python. This package ships mock / energy adapters and a WebSocket transport.

import { OrchestrationManifest } from "nexus-enterprise-agent";
import { RealtimeRuntime, RealtimeSession } from "nexus-enterprise-agent/realtime";
import { WebSocketTransport } from "nexus-enterprise-agent/realtime";

const manifest = await OrchestrationManifest.load("team.yaml");
const runtime = RealtimeRuntime.fromManifest(manifest, { runContext: ctx });
const pipeline = await runtime.buildPipeline("voice_agent");
const session = new RealtimeSession(pipeline, new WebSocketTransport(ws), sid);
await session.runAudio();

Learn more

Full docs: docs/index.md

| Topic | Doc | |-------|-----| | Architecture | docs/architecture.md | | YAML walkthrough | docs/getting-started.md | | TypeScript API | docs/getting-started-typescript.md | | Voice / channels | docs/reference/realtime-agents.md | | Pipelines | docs/guides/pipelines.md | | SaaS wiring | docs/guides/saas-example.md | | All examples | docs/examples.md |