@automatey-org/agent-core

The one framework-agnostic Vercel AI SDK ReAct streaming loop — shared by the Automatey CLI, the web app, and the browser extension. No React, no DOM, no provider construction. Just the turn loop, a stable callback contract, and types.

import { Agent } from '@automatey-org/agent-core';
import { createOpenAICompatible } from '@ai-sdk/openai-compatible';

const model = createOpenAICompatible({ name: 'vllm', baseURL, apiKey })('nvidia/…');

const agent = new Agent({
  model,
  tools,                       // Record<string, Tool> (Vercel AI SDK ToolSet)
  systemPrompt,
  maxSteps: 8,
  modelLabel: 'Nemotron',
  extractReasoningTag: 'think', // optional: pull <think>…</think> into reasoning
});

await agent.run('hello', {
  onToken: (t) => …,           // answer text delta
  onReasoning: (t) => …,       // live reasoning delta
  onToolCall: (name, input) => …,
  onToolResult: (name, output) => …,
  onStatus: (s) => …,          // 'thinking' | 'generating' | 'tool' | 'done'
  onMetrics: (m) => …,         // { model, durationMs, firstTokenMs, *Tokens }
  signal,                      // AbortSignal
});

Design

• Provider-agnostic: pass an already-built model. Provider construction (Nemotron thinking budget, OpenAI, Anthropic) lives in @automatey-org/nemotron.
• Minimal core, optional features: the <tag> reasoning extraction is opt-in (extractReasoningTag); everything else is the smallest viable loop.
• Vercel AI SDK v7: streamText + stopWhen: stepCountIs(n), unified fullStream (text-delta, reasoning-delta, tool-call, tool-result, finish-step). ai is a peer dependency so there is one SDK instance.

Peer dependency: ai >= 7.