Most assistants lose the thread after one response and treat an intended action as a completed one. I wanted a system that could hold onto a goal, operate across the desktop, recover when a step failed, and show what actually happened.
local-first AI · macOS
ORION
A persistent command hub for work on the Mac.
ORION keeps an objective alive across multiple steps, routes work through local tools and specialized workers, and checks evidence before it reports completion. The interface is a hub for supervising that process—not a chat window.
Based on Concept-Bytes/Jarvis ↗; substantially extended as ORION.
- Year
- 2025—Now
- My role
- Creator · agent architecture, macOS UI, memory, and safety
- Built with
- Python · Swift · SQLite · OpenAI
✓ SIGNAL 01 · INSPECT
Workspace and requested scope verified.
✓ SIGNAL 02 · ISOLATE
Only intended source files selected.
▶ ACTIVE VECTOR · VALIDATE
Running production checks and collecting evidence.
NEXT VECTOR
Report the verified result and any limitation.
I turned the original Jarvis foundation into a persistent macOS command hub with supervised goals, worker routing, voice interaction, local memory, computer actions, permission gates, recovery, and an execution trace designed around evidence rather than conversation bubbles.
ORION now spans 36 local modules and 210 tests. It can break a request into bounded work, keep the user in control of risky actions, and verify the result before closing the loop.
A desktop agent has to remain useful across long, branching work without hiding risk or confusing intention with completion.
I designed the agent architecture and macOS hub, then implemented persistent goals, memory, worker orchestration, permission controls, recovery behavior, execution traces, and a test suite covering the system’s tools and safety boundaries.
PythonSwiftSQLiteAgent architecturemacOS automationSafety systems
Technical detailsExplore the product systems+
What the product actually does.
Goals that survive more than one response
A persistent objective model tracks the active goal, current step, blockers, and recovery path instead of reducing the session to a stream of messages.
Local tools and specialized workers
Desktop inspection, memory, transcription, and fast actions stay local when possible; larger tasks route to bounded workers with explicit permissions.
Completion requires proof
Every action exposes its target, status, result, and failure state. ORION verifies observable evidence before it reports that the task is complete.
Continue exploring
Lyfora