AI Agent Framework (@subhajit60/aiagent)

A modular, extensible, and clean-architecture TypeScript framework designed to build autonomous agents equipped with planners, LLM providers, conversation memory, and custom tools.

Features

• Decoupled Architecture: Built around interfaces allowing you to easily swap any module (LLM Provider, Memory, Tools, Planner, etc.).
• Planners: Autonomous plan-generation support that decomposes high-level goals into step-by-step actions.
• LLM Providers: Ready-to-use providers for:
  • Google Gemini (via @google/genai)
  • OpenAI (Skeleton structure, easily expandable)
  • Ollama (Local models)
• Memory Systems:
  • ConversationMemory (In-memory conversational message stack)
  • VectorMemory (With keyword fallback search and mock vectors for semantic search integrations)
• Extensible Tools: Includes built-in support for:
  • File Operations (FileTool)
  • Command Execution (TerminalTool)
  • Browser Control (BrowserTool)
• Robust Execution Loop: Prevents infinite loops, gracefully processes unstructured LLM output, and parses tool/answer structure.

Installation

npm install @subhajit60/aiagent

Quick Start

Here is a simple example showing how to initialize an LLM Provider, a local Conversation Memory store, a custom tool, and execute the Agent loop:

import { 
  Agent, 
  GeminiProvider, 
  ConversationMemory, 
  ToolRegistry, 
  FileTool 
} from '@subhajit60/aiagent';

// 1. Initialize the LLM provider (mocked by default, or connects to external endpoints)
const provider = new GeminiProvider({
  apiKey: process.env.GEMINI_API_KEY
});

// 2. Initialize memory to track the conversation flow
const memory = new ConversationMemory();

// 3. Register tools that the agent can execute
const tools = new ToolRegistry([
  new FileTool()
]);

// 4. Instantiate the Agent
const agent = new Agent({
  name: 'DemoAgent',
  provider,
  memory,
  tools
});

// 5. Run the agent execution loop
const finalResult = await agent.run('List the files in the directory and summarize the project.');
console.log('Result:', finalResult);

Running the Built-in Demo

Clone this repository and run:

npm install
npm run build
npm start

This runs the example demo using a mock fallback for Google Gemini, allowing you to trace the execution cycle of reasoning and tool use.

License

MIT License. See LICENSE for more information.