Repository navigation
Make byob look like anything: shader backdrops with frosted glass, panels placed anywhere, new button styles. HTML apps can use the internet when you or byob let them. Plus a benchmark for the model you're connected to.
Cockpit (experimental)
Ask byob ("make it a starship's bridge") or use Settings → Appearance → Cockpit.
- Backdrops: a shader scene drawn behind byob, and the panels turn to glass over it. byob writes it for you in WGSL, for whatever place or mood you describe. A scene can react to the pointer, to the button under it, and to the model working on a reply. byob compiles a new scene before it runs and refuses it with the line and reason when it doesn't. If byob closes within a few seconds of a new backdrop starting, that backdrop is switched off at the next start, and
BYOB_NO_BACKDROP=1starts without one. - Frosted glass: the scene is blurred behind the panels and stays sharp between them, in the desktop app and the browser. It needs a backdrop on when byob starts.
- Free arrangement: the header, chat list, chat, message box and panes can be placed anywhere, each in a glass, instrument or bare frame. Frames can carry labels. Start from floating or focus, then drag, resize or hide parts with ⋯ → Arrange panels…. Dock everything goes back to the usual layout.
- Controls: standard, HUD, plate or pill buttons.
- From a terminal:
--backdrop FILE.wgsl|off,--arrangement NAME,--look KEY=VALUE.
HTML apps on the internet
- An HTML app can be given internet access: Internet: on in its pane, or byob turns it on itself when it makes an app that needs live data or a web API. It works on the desktop and in the browser.
- Access is HTTPS only (fetch, plus scripts, styles, fonts and images by URL). The app stays sandboxed and can't reach byob itself or plain-HTTP addresses. Apps stay offline unless they're given access.
Benchmark
/benchmeasures the chat's model: generated tokens per second in total and per request, prompt processing speed and time to the first token, at each number of parallel requests up to the server's slots, for several prompt sizes. It draws four charts and makes no model request. Options:/bench c=1,2,4 p=512,4k gen=256 runs=1 charts=off.- Asked to "benchmark yourself", the model runs the same measurements and explains the results. The bundled Benchmark plugin can be switched off.
Fixes
- A reply the server cut off at the model's output limit now says so, instead of ending as if finished or blank. This mostly affects models that think at length. Ask the model to go on, or raise Max tokens in the model's settings.
- byob reports GPU errors instead of closing on them.
- On a phone, Show artifacts and the floating Artifacts button open the newest artifact when none is open on that device yet. Before, an artifact made from the desktop couldn't be opened that way on the phone.
- On macOS, two spreadsheets saved at the same instant could overwrite each other's files.
The macOS builds are still not signed by an Apple developer; see the README for opening them (or fetch the archive from a terminal).