Ghostie

A desktop agent experiment in moving AI beyond conversation and into a person's computer and daily actions.

Product page · Download a historical release

The value of a personal assistant is not that it can always answer, but that it can gradually enter a person's environment of action.

A chat model can answer a question, yet a boundary remains between an answer and an action. Files live on the computer, tools belong to the operating system, tasks unfold through time, and each person develops their own preferences and ways of working. A personal agent has to cross that boundary: understand context, use local capabilities, and allow work to continue after the conversation ends.

Ghostie was an early desktop experiment around that idea.

Beginning with a Chat Window

Ghostie began as a desktop assistant that could be summoned at any time. Users could configure different models and assistants with distinct roles, then use one interface for everyday questions, writing, analysis, and programming work.

But more convenient chat was not the final objective. The significance of the desktop was proximity to the file system, local data, and system tools. That proximity created a path from generating answers toward executing tasks.

Composable Capabilities

Ghostie explored a system built from assistants, tools, MCP, knowledge bases, workflows, and schedules:

• assistants defined roles, context, and behavior;
• local tools and MCP connected models to practical capabilities;
• knowledge bases made personal material retrievable during a task;
• a visual workflow editor organized agents, tools, conditions, and execution paths;
• schedules explored work triggered by time or events.

Together, these modules supported one conclusion: a personal agent should not be a closed universal assistant. Its capabilities should be organized by the user around the way they actually live and work.

An Early Exploration

Ghostie took shape before I began studying agent products as a broader system. It explored multiple models, desktop execution, knowledge, tools, workflows, and scheduling, while also exposing the complexity of concentrating all those concerns inside one desktop application.

The project is now preserved as a historical release rather than an actively developed primary product. The questions it validated later became a clearer division of responsibility: Downcity addresses agent runtime and productization infrastructure, while Duobox addresses the project environment in which humans and agents work together.