Web-Agent-Interface Library

Overview

The web-agent-interface library provides tools to LLM agents in browsers to interact with different websites.

Note: The project is still in development phase. It has limited coverage and its APIs may change.

Supported Tools

• General DOM Operations: Get page content and user selection.
• Overleaf: Edit Overleaf documents.
• Google Calendar (GCal): Read and create events on Google Calendar.

Installation

To install and build the library, follow these steps:

1. Clone the repository:

   git clone https://github.com/mlc-ai/web-agent-interface.git
   cd web-agent-interface

2. Install dependencies and build the project:

   npm install
   npm run build

Usage

1. Initialize State

import { State } from '@mlc-ai/web-agent-interface';

const state = new State();

2. Import Tools

import { tool, retriever, action } from '@mlc-ai/web-agent-interface';

3. Give Tool Description to Prompt

const system_prompt = `
You are a helpful AI agent.

You have the following tools to use:

${tools.map((t) => JSON.stringify(t.schema)).join(",\n")}
`

3. Call Tool Function to Get Observation

const { tool_name, parameters } = extractToolCall(llm_response);
const observation = tool[tool_name].implementation(state, parameters);
console.log("Got observation:", observation);