A workstation-based AI worker for coding, analysis, documentation,
and real business operations.
WorkBox Labs explores how far local AI can go when it has the freedom to think —
but operates on hardware you own, and actions you can approve or reject in plain language.
Local-only workloads
No hidden network channels
Human-approved actions
Optical one-way output (GlassBox)
Projects
What WorkBox Labs is building
The long-term goal is simple: show that serious, high-value work can be done by AI
that lives on your own machines, with transparent limits and visible, reversible actions.
WorkBox AI
A workstation-based AI worker for coding, analysis, documentation, and real
business operations tasks — all on hardware you own.
Workstation
WorkBox AI behaves more like a focused software engineer and operations assistant
than a chatbot.
It streams its desktop, takes on structured tasks, and works end-to-end on code,
documents, prototypes or planning workflows — with every input and file optically
ingressed, sanitized, and under your control.
A research prototype that keeps AI behind glass with a single, visible output channel.
Research
The GlassBox Isolator runs a powerful local model in an Isolator with no network, no
keyboard and no mouse. The model can “think” freely, but it can only propose actions by
emitting a Human Accountability Ticket (HAT) as a QR code.
A separate console decodes each HAT, shows the full plan, and requires human approval
before anything runs on a separate machine. Every step is logged.
Today’s AI tends to live at two extremes: tightly sandboxed tools that can’t help much,
or highly-connected agents you can’t fully see. WorkBox Labs explores a middle path:
powerful local AI workers, wrapped in visible, human-first safety rails.
Local-first, not cloud-first. Your data, code and files stay on machines you control.
Visible actions. GlassBox uses HATs so every action is described in plain JSON and UI before it happens.
Human approval baked in. The console shows previews, tiers and audit trails so you decide what runs.
No hidden backchannels. The Isolator exposes only a one-way optical output — there is no inbound control path.
Designed for real work. WorkBox AI focuses on end-to-end tasks, not just conversation snippets.
Where things are up to
The current prototypes are already capable of generating HATs, displaying QR bursts,
decoding them on a main console and running approved actions with logging in place.
Next milestones include refining the WorkBox AI worker loop, polishing the GlassBox
demo site and shaping early, carefully-scoped pilot use cases.