@mast-ai/core

Core runtime for MAST, a TypeScript library that runs AI agent loops in the browser. The "think-act" loop, tool registry, and conversation state all live client-side; remote backends serve only as stateless reasoning engines.

Install

npm install @mast-ai/core

What's in the box

• AgentRunner — drives the think-act loop and emits an AgentEvent stream (text deltas, thinking, tool calls, tool results, done, error).
• ToolRegistry — browser-side tool store. Tools are plain TypeScript functions with direct access to the DOM, localStorage, and any client state.
• createAgent — config helper for agents (name, instructions, tool list).
• UrpAdapter + HttpTransport — connect to a Universal Reasoning Protocol (URP) backend over HTTP (JSON or SSE).
• LlmAdapter interface — swap in any inference source. See the sister packages @mast-ai/google-genai (direct browser-to-Gemini) and @mast-ai/built-in-ai (on-device via Chrome Prompt API).

Quick start (Hybrid Mode)

import { ToolRegistry, HttpTransport, UrpAdapter, AgentRunner, createAgent } from '@mast-ai/core';

const registry = new ToolRegistry().register({
  definition: () => ({
    name: 'getScreenResolution',
    description: "Returns the user's current screen width and height.",
    parameters: { type: 'object', properties: {}, required: [] },
  }),
  call: async () => ({ width: window.innerWidth, height: window.innerHeight }),
});

const agent = createAgent({
  name: 'BrowserAssistant',
  instructions: 'You are a helpful UI assistant. Use tools to answer questions about the screen.',
  tools: ['getScreenResolution'],
});

const transport = new HttpTransport({ url: 'http://localhost:3000/api/chat' });
const adapter = new UrpAdapter(transport);
const runner = new AgentRunner(adapter, registry);

const result = await runner.run(agent, 'How big is my screen?');
console.log(result.output);

Documentation

• Repository README
• Product Requirements
• URP Protocol Spec

License

Apache-2.0. Copyright 2026 Andre Cipriani Bandarra.