I started building Gideon because I believe AI agents will become a normal part of personal computing, but only if users can clearly understand and control what those agents are allowed to do.

Gideon is a Windows desktop agent built to help users complete real work through secure conversations, durable missions, reusable automations, inspectable memory, and scoped computer actions.

What makes Gideon different is that it was not built as a thin interface around a single model API. It is powered by Munesoft Agent, the TypeScript framework I built for model routing, orchestration, workflows, memory, approvals, plugins, tracing, evaluation, and recovery.

My goal is to make authority visible.

With Gideon, users can:

• Choose the workspace an agent may access
• Review proposed actions before they happen
• Approve the exact action or revision
• Receive a durable receipt of completed work
• Revoke access whenever they choose
• Use different AI providers instead of being locked into one model

I have taken Gideon from an agent framework to a packaged desktop product with isolated processes, schema-validated IPC, encrypted credential storage, crash recovery, approval-bound actions, and extensive automated testing.

It is still early, and I am especially interested in feedback on:

• Which tasks you would trust a desktop agent to handle
• Which actions should always require approval
• What would make you comfortable keeping an agent installed on your computer
• Which integrations and automations would be most valuable

Thank you for checking out Gideon. I’ll be here throughout the launch answering questions, hearing criticism, and learning from the community.