LoRA Studio

Runs free on your GPU box.

1. 01

   You open it, or the assistant does

   Opening ?app=lora loads its 1 screen into the shared cockpit, starting on lora-studio. The ribbon, the ticket list and the assistant are already there — the app contributes views, not a second application to log into.

2. 02

   Or you just ask

   The assistant classifies what you asked for and hands it to the bot that owns that domain. You do not pick the app, and you do not learn a command — routing is by declared capability, not by keyword matching a menu.

3. 03

   A named bot picks it up

   lora-director is accountable for the work. 4 declared capabilities say what it is trusted to do; the tokens and dollars it spends are recorded against that identity, so the cost has a line item rather than disappearing into a monthly bill.

4. 04

   Longer work becomes a ticket

   Scheduled and swarm-initiated work enters the queue as a lora-train ticket. It is durable, it has an owner, you can watch it move through its phases, and if it fails it fails somewhere you can open and read.

5. 05

   The result lands where you can check it

   Back on the app's own screens, as state you own — not as a chat message you have to scroll back to find.

• Routing is by declared capability, not by keyword. lora-director advertises what it is accountable for, and the assistant matches against that.
• Asks that land here: anything about lora training direction, anything about dataset curation, anything about validation analysis, anything about character consistency.
• The specialist does the work, not the assistant. Cost, permissions and audit stay attached to the bot above — delegation does not launder accountability.
• It will tell you when it cannot. If a request needs your screen — an approval, a payment, a login — it says which app to open rather than guessing.

All 11 creative apps →