Scrapalot Apps

• Mac dialogs now have the same built-in close control as Windows and Linux. The small ✕ appears in the top-right corner only while a dialog is open, without duplicating the macOS menu bar or shrinking the workspace
• Google sign-in now returns to the desktop app instead of stopping on the web dashboard. The browser securely resumes the one-time desktop authorization and hands the session back to Scrapalot
• One way out of every dialog, always in the same place: a small ✕ at the far right of the menu bar, there only while a dialog is open. Dialogs no longer draw a close button of their own, and the ✕ no longer squeezes in next to Help — the menu bar is now drawn by the app itself so it can hold it at the edge. File, Edit, View, Window and Help, their shortcuts and the Alt+letter keys all work as before
• Settings, Knowledge Stacks and the Data Inspector fill the window instead of opening as separate windows. A window of their own read as a second application: another frame to find, another copy of the app to load, and a menu bar above it whose "New Chat" belonged to the window behind
• Escape closes what is on top and nothing else. A dialog opened from inside another one — naming a new knowledge stack, the page-width menu over a PDF — used to take the panel underneath down with it, and whatever you had typed there went too
• Local AI is a running engine on all three systems in this build. Windows and macOS were a hardware panel in 0.7.8 because the machine that cut it was Linux; this one is built on each system, so the engine is the one that belongs there
• Speech-to-text works on Windows. The installer carried the Linux transcription runtime by mistake, so there was nothing for it to run
• The menu bar does the app's work: start a chat or a note, upload, open your library, jump to a settings page, search your conversations, show the sidebar. Help now answers what a support message asks for — versions, where the local engine is listening, and where your data lives, copied in one click
• Local AI lists the models on THIS computer. Models downloaded on another machine used to take their place — a single model from a Mac could hide the thirteen sitting on your own disk
• Books your library already holds are no longer offered again as new. On one mounted folder that was 58 files out of 89
• A mounted folder is recognised before its files have been fingerprinted, so a library this app already knows stops being treated as new
• The shelf says what the local engine is doing and what it already finished, instead of showing a completed batch as if it had never been touched
• Mac only, and a first cut: models built for Apple Silicon. Manage Models now lists MLX repositories on an M-series Mac instead of the GGUF ones, and a second engine runs them — it installs itself the first time you deploy one, so the first start takes a while and wants a working internet connection
• Speak instead of typing: transcription runs on this computer and the recording never leaves it — pick a model size once and it downloads in the background
• Every installer now carries the local-AI helper. 0.7.0 could ship without it, and the Local AI tab then quietly fell back to reading your hardware and nothing more
• An Apple Silicon Mac reports its own graphics again — it was calling itself CPU-only, because the check it went through only answers when PyTorch is installed and the helper deliberately leaves PyTorch out
• Delete books straight from a mounted folder — they go to the recycle bin, and nothing outside the folder you mounted can be touched
• Tick several books on this computer and remove them in one go
• A book already processed on this computer is marked as such instead of looking untouched (Windows picks this up with its next build)
• Covers appear before a book has been fingerprinted, not after
• The staged shelf shows the collection's folder rather than the whole mounted drive
• Compose counts the local books you ticked
• Local AI runs on Linux now: your own graphics cards, your own models, the same engine Windows uses — it arrives on first start, nothing to install
• The Linux download did not grow doing it — it had been carrying a Windows-only helper it could never run, and that is gone
• Every one of these reaches a Mac for the first time — the Mac download had been stuck four releases behind
• On a touchscreen, the ⋮ button on a saved memory is there without hovering first
• Mounting a folder no longer bogs the app down: no background fingerprinting, no re-walking the folder at every start — covers appear while you scroll
• Sorting asks the database, so "newest" means newest in the whole collection, not just the page you loaded — with a proper ascending/descending button
• Select books on your computer and process them in one batch, with the steps you chose
• Local embeddings use the same model the cloud does (all-MiniLM-L6-v2), downloaded once and run on your own GPU
• Books from a folder you mount show up instantly, the way opening a folder does — covers included (PDF and EPUB), and they open in the same readers as the rest of your library
• You decide what happens to each book: titles and content, embeddings, knowledge graph — each one on your own computer or in the cloud
• Local processing runs on your hardware: text extraction and full-text search stay on this machine, embeddings go through your own GPU
• Sorting covers local books too, with a proper ascending/descending button
• The library no longer refreshes itself every few seconds, and book covers load only for what you are actually looking at
• Chats with Ollama, LM Studio or vLLM on your computer now go through the full Scrapalot pipeline — your collections, citations and saved conversations all work, while the model itself runs on your machine
• First launch now asks how you want to start: sign in through your browser, or continue as a guest
• Model search in Manage Models is live — it searches HuggingFace and ranks results by what your graphics cards can actually run
• Local AI settings now describe YOUR computer — GPUs, memory and models on this machine, powered by a bundled local engine
• A machine with two identical graphics cards is finally counted as having two
• Scan Models scans — it no longer claims the feature is coming later
• When a signed-in session expires you are offered a way back in, instead of being quietly demoted to a fresh guest
• Ollama, LM Studio and vLLM running on your computer are found at localhost, models listed straight from them
• Signing in happens in your browser and brings you straight back — the app never shows a login form
• Already signed in? Switch accounts from Settings → Account with the same browser flow

• A model you downloaded now chats like Ollama does: your collections answer, citations come back, and the conversation is saved. Until now it replied and left nothing behind — no conversation in the sidebar, nothing to come back to
• LM Studio providers were being filtered out of the same lists for the same reason, and are not any more
• Local AI answers on a Mac. Chatting with a model you downloaded used to fail outright — the Apple Silicon engine went looking for the model on HuggingFace instead of using the copy already on your disk
• The Mac download finally carries the local-AI helper itself. Until now it was built on Linux, which cannot produce a Mac one, so the Local AI tab had nothing to talk to unless you installed a Python of your own
• Models on your computer sit at the top of the model list, right after Scrapalot AI, instead of below every provider you have connected
• A model you just downloaded shows up in the model list straight away — it used to need an app restart
• Choose which downloaded model to serve, from the Installed list, and see which one is running
• A local model answers in your language instead of picking one of its own
• The guided tour starts again, and its steps fit a phone screen
• The sidebar Go to list folds away instead of filling the screen
• The message box rises above the keyboard in team chats and rooms
• Tag a Stack or a document with @ or / while chatting to a team or room
• An update that downloaded but never applied no longer leaves the app stuck

Already have the app? It refreshes its own screens on the next launch. The ringtone is part of the app itself, so install this build to hear a call arrive.