Skip to main content
On this page

LosOS Tag

LosOS Tag is the assistant you tag in LosOS Chat. Write @LosOS in a message — or accept the suggestion the composer shows when a room's questions have gone unanswered — and the box answers in the thread with what its own search found: your files in LosOS Cloud, the code in LosOS Git, your mail, the rooms of chat. Nothing leaves the box; the answer runs on the box's own model.

Tag lives in Settings › LosOS Tag, where you shape four things.

Levels and areas​

A level is one of the box's words for how private a piece of data is, least private first (everyday by default; an administrator adds more — personal, secret, whatever the box's own words are). An area grades one piece of the box's data with one level: a folder in your files, a repository, a room, a mailbox, a journal unit.

The rule is one line: a tag reads an area only when its clearance reaches the area's level. The more private an area, the higher the clearance the assistant needs. Your own clearance (what the tag and the companion in your AI tools read as you) is yours to set on the same pane.

Assistants​

Every assistant is a real chat participant — @LosOS is the built-in one — with a persona of its own, a clearance, and a maker. Tag one in a message and it answers; tag two and they answer, and an answer that tags another assistant hands the turn on, three handoffs at most, so two assistants can carry a thread between them without looping forever. An assistant reads only up to its own clearance, whatever its persona says: the levels are the box's rule, not the model's mood.

Skills​

A skill is an instruction an ask can carry — the seeded ones come with the box (Find where I wrote it, Summarise this thread, What did I miss, Plan it), and you may write your own. A skill never raises a clearance and never grants a tool; it only says how to go about the answer.

What the tag never does​

Whatever its clearance, a tag never calls sudo or a shell, never buys, sells, rents or lends storage or compute, and never touches anything on disk in the owner's directories — their files are read and written through LosOS Cloud's own application, not by an assistant. A web search, when the box runs one (its own SearXNG, never the open internet's), takes the subject alone: no names, files or mail from the box ever go into a query.

Bigger models, over the mesh​

The tag's model is the box's one unified model (Bonsai 27B by default — a 1-bit recording of Qwen3.6-27B: 4.4 GB, tools, thinking, vision). A box too small to hold it asks over the network instead: in a mesh, the edge runs a head per model group and the boxes lend their memory to it (the mesh, AI models), so a 4 GiB box can ask Qwen 3.8 Max or DeepSeek R1 — the head answers, the lenders hold the weights. The ask queue pauses what nobody waits on while the box is short of memory, and resumes it on its own; nothing is dropped.