feder-cr/invisible_dots

What it solves

This project solves the problem of giving an AI agent a real, persistent computer to work on, where it can browse the web, log into sites, run programs, and perform long-running tasks without being blocked by anti-bot systems. It is an open-source, self-hosted alternative to commercial products like OpenAI Dots, Meta Muse, and Grok Bot, intended for people who want full control and privacy.

How it works

Each "Dot" (agent) runs inside its own hardware-accelerated virtual machine with a Linux desktop and a permanent disk. The VM sleeps when idle and wakes up for messages, tasks, or scheduled automations. The browser is a custom-patched Firefox (the engine is modified in C++, with fingerprints set inside the engine) so it looks like a real human browser and is not blocked. Every Dot keeps its own cookies and logins. The agent writes notes and how-tos in its home folder to serve as memory and skills for later tasks. Users can attach their own MCP servers, set fine-grained permissions (allow, ask, or deny each tool), and manage a task queue with priorities and recurring jobs. Interaction happens through a web UI, CLI, HTTP API, Telegram, or optionally WhatsApp.

Who it’s for

The project is aimed at developers, power users, and anyone who wants a self-hosted AI assistant that can autonomously operate a computer, handle tasks like downloading invoices, checking flight prices, or managing web logins, while keeping full control and privacy. It requires Node 24+, Go 1.25+, Git, and hardware virtualization, plus an OpenRouter API key.

Highlights

• Each Dot runs in its own hardware-accelerated VM with a persistent disk, sleeping and waking on demand
• The browser (patched Firefox) is engineered to be undetectable by anti-bot systems, with per-identity cookies and logins
• The agent maintains its own memory and skills by writing notes in its home folder, similar to Claude Code
• Declarative MCP server support (both local programs and URLs) with keys kept out of config
• Granular permission system: allow, ask, or deny for every tool, with requests surviving restarts
• Task queue with priorities, start times, and recurring automations that boot the VM on schedule
• Multiple ways to reach it: web UI, CLI, HTTP API with live events, Telegram, and opt-in WhatsApp
• Crash-safe: messages and tasks survive restarts, and interrupted tool calls are never run twice