Build a tron with a conversation
Describe what you need and the tron builder designs it with you: agents, models, memory and tools, checked against the catalogue, proposed as a tron you import with one click. It runs on your own keys.
What it does
Build a tron (New tron → Build it with a conversation, or the Home) is a chat. The builder asks what it needs to know, picks an architecture (a single assistant, a head with specialists, research with verification, triage, a scheduled round), chooses models among those you have a key for, searches the catalogue for plugins, skills and MCP servers by their real ids, and checks the .tron document exactly as an import would, without creating anything. When it is ready it proposes: a card with the agents, their models and roles, the memories, the tools and the slots you will fill (keys, tokens).
Import, then test
IMPORT THIS TRON asks for a confirmation and imports in strict mode: if the proposal names something your catalogue does not have, nothing is created. The tron is born switched off in your Garage, with its slots listed. After the import you can let the builder test it: with your consent it switches the tron on, sends it one message, reads the answer and switches it off (or leaves it running, if you ask). One test per consent. From then on the builder works on that tron: its next proposals show UPDATE THIS TRON, which changes it in place (strict, like the import; agents kept by name with their chats and memories) and says how many things changed, beside IMPORT AS A NEW TRON; if the tron is deleted, the conversation says so and goes back to importing.
What it costs
Nothing from a plan: it runs on your keys. The conversation uses the model and the key you choose in Settings, at the foot of the conversations list (yours, or your company's); a short line under the message box says what is chosen and opens the same panel. JEV decides is optional and sits in the same panel: it runs on a key of your vault (yours or the company's), TypeSafe keys listed first; with any other key, a refusal from TypeSafe is shown as it is. JEV System One answers the builder's decisions (which architecture, which model tier, is the brief complete, is the proposal ready) with calibrated probabilities; without it the chat model answers them in JSON, one extra call per turn. Every call is in your usage ledger: Analytics shows them under Tron builder (purposes tron_builder and jev). A test also costs the tron's own turn, on its keys.
The turn limit
In one turn the builder may take at most 8, 16 or 32 tool rounds (searching, opening a card, validating, proposing): you choose under Turn limit in Settings, 8 to start with. When it reaches the limit it stops, says what it was doing last, and waits: say continue to go on, or raise the limit.
What it remembers
The builder keeps a short memory of you (at most 4 KB): your sector, the models you prefer, how you like to work. It rewrites it as it learns and reads it at the start of every conversation, so you can build several trons over several conversations. You can read, correct or clear it in Settings, under What it remembers about you. Conversations, proposals and the memory are encrypted with your key and seen by nobody else, not even your company's admins.