M. Arslan

Why Ghosthread exists

Every automation tool for Windows forces the same compromise: either brittle script-based macros that break every time a window moves, or cloud services that stream your files, screens, and clipboard to someone else's servers.

Ghosthread was designed to end that compromise. The engine operates apps through their own message queues, at the operating-system level, without ever touching your cursor or focus. It runs on your machine and, by default, stays there. Every architectural decision from the first line was made to preserve that single promise.

How it was designed

Ghosthread is not a wrapper around an LLM. It is an execution substrate that an LLM plugs into. The design starts with a typed contract layer — every action declares what it consumes and produces, and the executor validates those declarations before anything touches the operating system. Safety is enforced below the AI, at the boundary where the AI no longer has a say.

On top of that sits a Directed Acyclic Graph scheduler that decomposes complex requests into ordered, isolated steps, each running in its own lane — IO, CPU, GUI, or System — so a hung GUI action cannot starve a file operation. Then a perception stack that reads screens directly rather than relying on app cooperation. Then a safety model with five independent layers, each one refusing to trust the one above it.

The architecture is documented on the How It Works page. The security model is on the Security page. Both reflect the implementation as it stands — not aspirational design docs.

The ten-phase build plan

Ghosthread ships when the vision is complete, not on a date someone else picked. The build is structured as ten independently-usable phases — reliability first, then Prism, then app controllers, then the agent loop, then the memory graph, then the domain specialists, then extensions, then voice. Each phase ships something a user can touch. Nothing is held back for a big reveal.

The roadmap page documents every phase in full. Nothing on it is aspirational — every phase is committed, every phase has a defined scope, and the sequence is fixed even if the calendar isn't.

Background

Formal training in computer science, combined with independent research in offensive security, penetration testing, and defensive SOC design. That security background is why Ghosthread ships with threat interdiction, path jails, DLP filtering, and rejection of destructive commands built in — before any AI capability was added. The safety model was not bolted on after the fact. It was the first thing designed.