Documentation

Interface and conversations

Each LlamaBoss window is its own chat workspace. Windows that use a local model share one model service, so the model is only loaded into VRAM once.

Applies to LlamaBoss 0.1.23 for Windows. Updated October 12, 2026.

Top bar

ControlWhat it does
Model pillShows the current model. Click it to switch between local GGUF models and models from AI providers you've connected.
Thinking chipThe · Auto label inside the pill. Click it to choose Auto, Off, On, Low, Medium or High thinking for this chat. It applies from your next message. See Thinking.
Robot buttonTurns Agent Mode on or off for the current chat.
Project and SkillsAttach or manage a Project, and create, open, import or export Skills. The former Goal control was retired.
Context meterShows ctx used/total for the next request. The number is exact after each completed reply; a ~ means it's an estimate. Amber means older tool output is being trimmed from what the model sees, and red means the window is nearly full, so consider a new chat. Click the meter to open detailed context and response statistics. Turn it on in Settings → Behavior.

Open the sidebar with the ☰ button. Conversations are grouped under their Project or Unassigned. Each group has Pinned, Today, Yesterday, Previous 7 days and Older sections, and can be collapsed. Dates follow chat activity; reopening a chat alone does not move it to Today.

  • Search: press Ctrl+F or Ctrl+K, type, then use the arrow keys and Enter to open a result. Esc clears the search.
  • Keyboard: with the list focused, the arrow keys move between chats.
  • Actions: each row's menu can rename, pin, archive, export, delete, or move the chat to a Project. You can also drag a chat onto a Project.
  • Several at once: Ctrl+click or Shift+click to select, then pin, archive, move or delete them together.
  • Archive: archived chats leave the main list. Open Archived in the sidebar to browse or restore them.
  • Export: export one chat from its menu as Markdown or text. Export is a readable transcript; save/open uses the conversation JSON.

Activity line

The thin line above the composer shows what the agent is doing right now, such as which tool is running and for how long, or that it's waiting for your approval. If a tool runs for more than about 20 seconds, the taskbar button flashes when it finishes so you can work in another window meanwhile. Very long tools offer a Stop Long-Running Task option. From 0.1.23, the line appears only once a tool has run for about 0.75 seconds, so quick tools no longer make the chat jump.

Images and hashes in answers

From 0.1.23, an image written in Markdown on a line of its own, such as ![chart](C:\Reports\chart.png), appears as a thumbnail when it points to a PNG, JPEG, GIF, BMP, ICO or TIFF file on your PC. Images on the web are not downloaded; a short note with the link is shown instead.

Also from 0.1.23, after an agent reply LlamaBoss checks long hexadecimal values in it, such as SHA-256 hashes (32 characters or more). If one does not appear in this chat's tool results or your own messages, a notice under the reply points it out, since a model can garble a hash when copying it. The notice is shown only to you and is not sent to the model.

Multiple windows, one local model

Click + New Window in the sidebar, press Ctrl+Shift+N, or launch LlamaBoss again while it's running. Every window is a full chat workspace with its own conversation, Project, attachments and Agent Mode state.

Local-model windows share the running model. If one window is generating, a message from another waits its turn instead of interrupting. Windows using an AI provider don't touch the local model at all, and each can use a different provider model at the same time.

Composer

  • Enter sends and Shift+Enter adds a new line. You can drag to make the box taller.
  • Attach files with the paperclip, drag and drop, or paste an image.
  • Large pasted text becomes a text attachment instead of filling the box. Ctrl+Shift+V pastes it straight into the message.
  • While the model is answering, Send becomes Stop, which cancels the reply, a running tool sequence, or a pending approval.
  • If the model is still loading, your message waits and sends when it's ready.

Saving and loading

Conversations save automatically as messages, tool results and metadata change. Background saves keep file writing out of the main chat flow. Before leaving a chat or closing the app, LlamaBoss completes a final save. If that fails, you can retry, save a recovery copy elsewhere, or keep the chat open. You can also use:

ShortcutAction
Ctrl+NNew chat in this window.
Ctrl+Shift+NNew window.
Ctrl+SSave the conversation.
Ctrl+OOpen a saved conversation.

A conversation already open in another window is brought to the front rather than loaded twice. Deleting a conversation also removes its chat folder and the generated files and attachments inside it; export or copy anything you want to keep first.

Export with metrics

Open a conversation row's menu in the sidebar and choose Export with metrics.... The export includes the readable transcript and a metrics section drawn from the chat's recorded reply and context logs.

Where recorded, metrics include token usage, cached and reasoning tokens, time to the first token, total response time, and request/context composition. Missing measurements are shown as unavailable rather than invented. Older conversations may have only partial metrics. This is a usage and performance record, not a provider billing statement.

The underlying logs are turn_stats.tsv and ctx_calibration.tsv in the chat folder. Review the transcript and metadata before sharing an export; it can contain your messages, tool results, and personal paths.