AIKit - Extensible AI Assistant for Firefox

Note: This is a Proof of Concept (PoC) demonstrating AI-powered browser automation through an extensible tool system. Built with vibecoding - following the flow and exploring what's possible.

AIKit is a modular add-on system for Firefox that enables interaction with various AI models (Anthropic, OpenAI, Google) through a sidebar interface, using an extensible tool system.

⚠️ Security and Privacy Warning

This project is a security and privacy nightmare by design. Before using AIKit, understand what you're sharing:

What Gets Transmitted to Third Parties

When you use AIKit, the following data is sent to AI providers (Anthropic, OpenAI, or Google):

• Your prompts and commands - Everything you type in the chat
• Tab information - URLs, titles, and metadata of all your open tabs
• Browsing history - When using history search
• Bookmark data - All your saved bookmarks with URLs and titles
• Page content - Full HTML, text, and DOM structure when using DOM tools
• Screenshots - Visual content of web pages
• Download information - Files you're downloading, their URLs and names
• Container data - Your Firefox container setup and organization

Privacy Risks

• Data aggregation: AI providers could theoretically build detailed profiles of your browsing behavior
• Third-party storage: Your data passes through and may be stored by big tech companies
• No control: Once data is sent, you have no control over how it's processed or stored
• Potential logging: Conversations and tool calls might be logged for AI training or monitoring
• Sensitive information: Personal data, credentials, financial info visible in tabs/history could be exposed
• Cross-site tracking: Your browsing patterns across different sites become visible to a single entity

Recommendations

• DO NOT use with sensitive data: Banking, healthcare, personal communications
• DO NOT use on work computers: Company data could be exposed
• Use a separate browser profile: Create a dedicated Firefox profile for AIKit
• Review what you share: Be conscious of what tabs are open and what you're asking
• Consider self-hosted alternatives: Look into local AI models if privacy is critical
• Read provider policies: Understand Anthropic/OpenAI/Google's data retention policies
• This is a PoC: Not intended for production use or sensitive environments

Why This Architecture?

This design prioritizes functionality and ease of development over privacy. The AI needs context to be useful, but this comes at a significant privacy cost. Future iterations could explore:

• Local AI models (though limited capabilities)
• Differential privacy techniques
• User-controlled data filtering
• On-device processing where possible

Use AIKit knowing that you're trading privacy for convenience. Consider the risks carefully.

Architecture

The project consists of:

aikit-orchestrator (Main Add-on)

The main orchestrator that:

• Provides a sidebar with chat interface
• Integrates various AI providers (Anthropic, OpenAI, Google)
• Manages tool registration and execution from other add-ons
• Allows compatible add-ons to register and provide functionality

aikit-common (Shared Library)

Shared library that simplifies tool development:

• ToolProvider class for automatic tool registration
• Type definitions and interfaces
• Message handling and error management

Tool Add-ons

Available tools:

• aikit-tabs-tool - Tab management (open, close, list, activate, reload, search) and container management
• aikit-bookmarks-tool - Bookmark management (create, search, list, remove, organize)
• aikit-dom-tool - DOM manipulation (execute JavaScript, query elements, get page info)
• aikit-downloads-tool - Download management (list, pause, resume, cancel)
• aikit-history-tool - Browser history (search, recent items)
• aikit-screenshot-tool - Screenshot capture (visible area, full page, download)

How It Works

1. Tool add-ons register themselves with the orchestrator on startup
2. When the user makes a request, the AI analyzes the prompt and decides which tools to use
3. The orchestrator sends execution requests to the appropriate tool add-ons
4. Tool add-ons execute actions and return results
5. The AI uses the results to complete the response

Setup and Build

Prerequisites

• Node.js 18+
• npm

Build

./build.sh

This will build aikit-common, aikit-orchestrator, and all tool add-ons.

Other available commands:

• ./build.sh install - Install all dependencies
• ./build.sh clean - Clean build artifacts
• ./build.sh build --addon aikit-tabs-tool - Build specific addon
• ./build.sh help - Show all options

Installation in Firefox

1. Open Firefox and navigate to about:debugging#/runtime/this-firefox
2. Click "Load Temporary Add-on"
3. Select aikit-orchestrator/dist/manifest.json
4. Load each tool's dist/manifest.json you want to use

Configuration

1. Open the AIKit sidebar (View → Sidebar → AIKit Assistant)
2. Click the ⚙️ icon to open settings
3. Enter your API key for the chosen provider (Anthropic, OpenAI, or Google)
4. Save and initialize

Developing New Tool Add-ons

The beauty of AIKit's architecture is that the sky's the limit - you can create tools for virtually anything your browser can access. The current tools are just the beginning.

Potential Tool Ideas

The extensible system opens up endless possibilities:

Communication & Social

• Gmail integration (read, compose, search emails)
• WhatsApp Web automation (send messages, read conversations)
• Slack, Discord, Teams integration
• Twitter/X posting and timeline management
• LinkedIn automation

Productivity

• Google Calendar/Drive integration
• Notion, Trello, Asana task management
• GitHub/GitLab issue and PR management
• Jira ticket automation
• Cloud storage operations (Dropbox, OneDrive)

E-commerce & Finance

• Amazon shopping and order tracking
• Banking operations (balance checks, transaction history)
• Stock market monitoring
• Price comparison across sites
• Shopping cart automation

Content & Media

• YouTube video management
• Spotify/music control
• Netflix/streaming automation
• RSS feed aggregation
• PDF manipulation and reading

Development Tools

• Code execution and testing
• API testing and monitoring
• Database queries
• Server management
• DevOps automation

And Much More... The only limit is your imagination! Any web service or browser capability can become an AI-controlled tool.

Getting Started

See CONTRIBUTING.md for detailed instructions on creating new tool add-ons.

Quick overview:

• Use aikit-common library for simplified implementation
• Define tools using TypeBox schemas
• Implement handlers for your tools
• Use ToolProvider class to handle registration automatically

Example:

import { Type } from "@sinclair/typebox";
import { ToolProvider, ToolDescriptor } from "aikit-common";

const tools: ToolDescriptor[] = [
  {
    name: "my_tool.action",
    label: "My Action",
    description: "Description of what this tool does",
    parameters: Type.Object({
      param1: Type.String({ description: "Parameter description" })
    })
  }
];

const toolHandlers = {
  "my_tool.action": async (params) => {
    return {
      content: [{ type: "text", text: "Result..." }]
    };
  }
};

const toolProvider = new ToolProvider("aikit-my-tool", tools);
toolProvider.registerHandlers(toolHandlers);
toolProvider.start();

Supported AI Providers

Anthropic Claude

• Models: Visit docs.anthropic.com for current model list
• Features: Full tool calling support, streaming responses
• API Key: Get from console.anthropic.com

OpenAI

• Models: Visit platform.openai.com/docs/models for current model list
• Features: Full tool calling support with function calling API
• API Key: Get from platform.openai.com

Google Gemini

• Models: Visit ai.google.dev/models for current model list
• Features: Basic chat (limited tool support in current version)
• API Key: Get from makersuite.google.com

License

MPL-2.0

Copyright (c) 2026 Andrea Marchesini

Credits

Built with browser-compatible SDKs from:

• @anthropic-ai/sdk
• openai
• @google/generative-ai

Originally inspired by pi-mono by @mariozechner.