LamaDeck 0.9.9 — formerly Ollama Manager

The app has a new name: LamaDeck. "Ollama Manager" could suggest an official Ollama app — LamaDeck is an independent control panel for your local AI stack (Ollama, LiteLLM, Open WebUI and a built-in chat), not affiliated with Ollama. The repository moved to cyryllo/LamaDeck; old links and clones keep working.

Updating keeps everything. On its first start LamaDeck moves your settings (hosts, aliases, language), saved conversations, the LiteLLM router config and the start-after-login entry from the old "Ollama Manager" locations. install.sh replaces the old install by itself, and the .deb package replaces the old ollama-manager package.

✨ New in the chat

• Conversations are saved automatically after every answer — one file each in ~/.local/share/LamaDeck/czaty/, only on your computer. A list on the left opens and searches them (titles and content); right-click to rename, export to Markdown or delete (also double-click / F2 to rename, Delete to delete)
• Links in a question: paste a link and the app downloads the page itself — no account — and adds its text to the question (PDF links too; long pages go through excerpt search)
• Context usage: "4 051 / 65 536 (6%)" for the current conversation, with a warning from 80% — before the model silently starts forgetting the beginning
• Thinking on/off for models that reason first (qwen3, gemma4…): off, a simple question takes a fraction of a second instead of 15 s (measured on gemma4); reasoning can still be shown when on
• Small icons in message headers: copy an answer, regenerate the last answer with the model selected right now (handy for comparing models), edit the last question
• Model tags under the model picker: in memory / on disk, pictures, thinking, tools
• A more compact right panel and a wider conversation list
• The tray icon menu's "Open chat" opens the built-in chat

🔧 Install / update

• .deb: download lamadeck_0.9.9_all.deb from the Assets below and run sudo apt install ./lamadeck_0.9.9_all.deb — it replaces ollama-manager automatically. Run it as lamadeck or from the menu (Utilities).
• install.sh: git clone https://github.com/cyryllo/LamaDeck.git && cd LamaDeck && ./install.sh. If the app is already installed from a .deb, it now offers to remove that first, so you don't end up with two copies.

What's next

Before 1.0: an own APT repository (install and update with apt) and an AppImage for other distributions. 1.1 brings web search in the chat. After that: bug fixes only.

Ollama Manager 0.9.0

The last feature release before 1.0. The built-in chat can now read documents — PDF, Word, Excel, PowerPoint, LibreOffice and EPUB — with nothing beyond Ollama, Python and Qt.

✨ New

• Documents in the chat: attach .pdf, .docx, .xlsx, .pptx, .odt, .ods, .odp or .epub
  • the text is extracted in the background: tables and sheets as a; b; c rows, numbered slides and PDF pages, EPUB chapters in reading order
  • small documents are sent whole; large ones (100 KB – 20 MB of text) go through excerpt search with an embedding model in Ollama, and the answer shows which lines the excerpts came from
  • scanned PDFs without a text layer get a hint (paste a screenshot for a vision model, e.g. glm-ocr); old .doc / .xls / .ppt / .rtf files get a "save as" hint
• Estimated remaining time on the attachment while a large file is being indexed, based on the speed measured on your computer

🐛 Fixes

• Chat statistics card: descriptions wrap instead of being cut off by the values

📦 Packaging

• PDF reading uses Qt's own QtPdf module: it's already in the pip PyQt6; the .deb now depends on python3-pyqt6.qtpdf, and install.sh offers it when PyQt6 comes from apt without it

📖 Docs

• New README screenshots (dashboard, chat, tray panel) in English and Polish

What's next

0.9.0 is the near-final version. Before 1.0 only a few chat additions are planned (automatic saving of conversations, a context-usage indicator, a "no thinking" switch, regenerate/edit the last question, copying and exporting answers). After 1.0 the project switches to bug fixes only.

Installation

Choose one of the two ways — both work on Debian / Ubuntu / KDE neon and other Debian-based systems.

Option 1 — the .deb package (easiest)

Download ollama-manager_0.9.0_all.deb from the Assets below and install it:

sudo apt install ./ollama-manager_0.9.0_all.deb

apt pulls in all dependencies (PyQt6 with QtSvg and QtPdf, requests, pkexec, curl). The app appears in the menu under Utilities and can also be started with ollama-manager. Updating = installing a newer .deb the same way; removing = sudo apt remove ollama-manager.

Option 2 — install.sh (no root, installs into your home folder)

git clone https://github.com/cyryllo/Ollama-manager.git
cd Ollama-manager
git checkout v0.9.0
./install.sh

Installs PyQt6 and requests with pip for your user, copies the app to ~/.local/share/ollama-manager and adds a menu entry. Run it again to update (it detects the installed version and asks); remove with ./install.sh --uninstall.

Use only one of the two methods at a time — if you switch, uninstall the other one first, otherwise you'll get two menu entries.

After the first start, the app installs Ollama, the LiteLLM router and Open WebUI itself with one click each, if you want them.

Ollama Manager 0.8.6

The biggest update since the redesign: a built-in chat with speed stats, file attachments and search in large text files, plus a system tray panel, desktop notifications and a fix for random crashes.

✨ New

Chat (new screen)

• Chat with any model on any host, straight from Ollama — no Open WebUI or router needed
• Answers stream in with formatting (code blocks, lists, tables), a Copy code link above every code block, and a Stop button
• Speed stats for every answer, measured by Ollama: generation and prompt processing in tokens/s, token counts, model loading time, total time
• A notice with a seconds counter while the model is being loaded into memory (the first question can take a while)
• Optional view of the model's reasoning (thinking models such as qwen3)
• Attachments: 📎 button, drag & drop, or paste a screenshot with Ctrl+V
  • pictures go to models that can see images (vision); a model without vision is blocked with a hint
  • text files (code, .txt, .md, .csv, .json…) are added to the message
• Large text files (100 KB – 20 MB) are searched, not pasted: the file is split into excerpts and indexed with an embedding model in Ollama. Each question gets the best-matching excerpts, and the answer shows which lines they came from. Nothing beyond Ollama is needed, and nothing is written to disk
• The embedding model can be chosen in Settings → Chat. Without one, a button there downloads the recommended multilingual nomic-embed-text-v2-moe
• Only models that can chat are listed; OCR models appear only when a picture is attached

System tray (optional, off by default)

• Tray panel: the installed services with on/off switches, CPU / VRAM / RAM, models in memory, Open chat, plus a small WebUI button while Open WebUI is running
• Closing the window hides it to the tray; optional start after login; only one copy of the program runs at a time
• Works on Wayland (KDE Plasma): the panel opens next to the tray without flicker

Other

• Desktop notifications when a model download, installation, update or backup finishes while the window is in the background
• MoE filter in model search — handy for APUs and computers without a strong GPU
• RAM usage on the Ollama hosts screen
• Services that are not installed are hidden from the sidebar, Settings and the tray panel (installation stays on the dashboard)
• Router: the model list is now a table; a separator in the sidebar

🐛 Fixes

• Random crash ("the program closed by itself"): a worker thread could be destroyed while still finishing. All background threads are now kept alive until they really end
• Sidebar no longer jumps back to the dashboard when the window regains focus (Alt+Tab, Settings from the tray)
• Dark lines under updated values with fractional display scaling
• Strategy cards on the Router screen no longer get squeezed in longer translations

⚙️ Changed

• "Apply recommended values" in Settings now sets OLLAMA_MAX_LOADED_MODELS=2, so the chat model and the embedding model can stay in memory together

Install / update

./install.sh

Ollama Manager 0.6.8 — a whole new interface

The app has been rebuilt from the ground up: a dark Breeze look, a sidebar with seven
screens, bundled fonts (Inter + JetBrains Mono) and a new llama logo.

New screens

• Dashboard — one switch for all services (Ollama, LiteLLM router, Open WebUI), a card
  per service with its own on/off switch or an "Install" button, loaded models, GPU load,
  and live CPU / VRAM / RAM / network gauges.
• Ollama hosts — every host on your LAN at a glance: online/offline, Ollama version,
  response time, VRAM use and the models on each host (a RAM icon shows which ones are
  loaded). Add a host by typing just its address.
• LiteLLM router — a live route diagram, model aliases (fixed names like chat or
  code for a model on chosen hosts), a choice of what to do when a model is on several
  hosts (take turns / least busy / fastest), an optional API key, and ready-made settings
  for Continue.dev and Open WebUI.
• Models — models from all hosts as tiles with filters (chat, code, embedding,
  vision); Load / Unload, Copy to host, delete from one or all hosts. The Ollama +
  Hugging Face search stays, and downloads show speed and can be cancelled.
• Open WebUI — start / restart / stop, port, open the browser after start, require
  login, and a one-click backup of all chats (ZIP).
• Log — messages from Ollama, LiteLLM, Open WebUI and the app in one place, with a
  plain explanation of common errors; save to file.
• Settings — language, autostart and update checks for all three services, and
  Ollama environment variables, each with a short description.

Other changes

• The installer can give your account read access to the Ollama service log
  (one question, needs sudo; takes effect after logging in again).
• The "Help" tab and the "Manage servers" dialog are gone — descriptions now live in
  Settings, hosts are managed on the Ollama hosts screen.

Upgrading

• ./install.sh (it updates an existing install), or build a package with ./build-deb.sh.
• The .deb package now also depends on python3-pyqt6.qtsvg.
• The window is always dark, regardless of the KDE color scheme.

Ollama Manager 0.6.4

New

• Open WebUI connection guide. A new "Connection for Open WebUI..." button in the
  Model aggregator tab shows exactly how to add the LiteLLM aggregator in Open WebUI
  (Admin Panel → Settings → Connections → OpenAI API → "+"). It includes the address for
  this computer (http://localhost:4000/v1) and its LAN address (for WebUI running on
  another machine), with copy buttons. You only need to add one connection: WebUI then
  sees models from every host on the server list. When you add a new host, click
  "Restart" next to LiteLLM and refresh the WebUI page.
• Choose which models LiteLLM serves. Every model/host row in Models in the
  aggregator now has a checkbox. Unchecked models are left out of the LiteLLM config
  and the generated Continue config. Your choice is remembered after a refresh or
  restart. New models are on by default. The choice applies per host, so the same model
  can be served from one machine only (e.g. BC-250). Changes take effect after
  "Restart".

Removed

• The Unsloth Studio connection dialog.

Notes

• Earlier versions never connected Open WebUI to LiteLLM. WebUI only saw the local
  Ollama, so models from other hosts didn't show up there. The app shows instructions
  instead of writing the setting itself: Open WebUI keeps connections in its own
  database and ignores environment variables once a value is saved there.

Install / update: ./install.sh or build a package with ./build-deb.sh.

Ollama Manager v0.6.2

New

• Restrict LAN access by IP (IPAddressAllow): new field in the Advanced tab, right below OLLAMA_HOST.
  Enter the hosts allowed to connect, e.g. localhost 192.168.8.20 192.168.8.0/24.
  • Empty field = the whole LAN can connect (same as before).
  • With a list set, other private addresses (10.x, 172.16–31.x, 192.168.x) are blocked by systemd
    (IPAddressDeny), while internet access stays open, so model downloads keep working.
  • Entries are validated before saving (IP, CIDR subnet, or localhost / link-local / multicast / any).
  • ⚠️ The filter applies in both directions. If your DNS server is a router on the LAN (e.g. 192.168.8.1),
    add it to the list.
  • Only makes sense with OLLAMA_HOST enabled.
• Context length presets: OLLAMA_CONTEXT_LENGTH is now a drop-down with standard sizes:
  default, 4k, 8k, 16k, 32k, 64k, 128k.
  • You can still type any custom number, and non-standard values already in override.conf are kept.
  • Values that aren't a plain number are rejected before saving.

Other

• Translations updated for all UI languages (EN, DE, ES, FR, IT, PT).
• README updated with the new IPAddressAllow option.

v0.6.0

One model search for two sources

• The Download new model card is now a single search across the official Ollama library and Hugging Face (GGUF). Every result shows its Source; Ollama results are always listed first. Pick a model to see all its variants with real file sizes, then Download.
• Pasting a Hugging Face link jumps straight to that repo's variants; an exact name with a tag (e.g. llama3.2:3b) downloads without searching.
• Removed the recommended-models list, the Size/Quantization fields and the separate Hugging Face dialog. MLX-only and cloud-only models are skipped.
• The installed models table has a Source column (Ollama / Hugging Face).

Memory

• New RAM and VRAM + RAM tiles in the stats bar. For the local Ollama, RAM is measured from its processes (including memory-mapped model files); loaded models show their split, e.g. 20.0 GB VRAM + 16.0 GB RAM.
• The memory estimate for a download now uses the real file size.

LiteLLM aggregator

• Fixed: LiteLLM 1.105+ refused to start without a master key (fresh installs). New optional Require API key setting, off by default.
• Connection details for Unsloth Studio (it has no config file).

Fixes

• Ollama search broke after ollama.com changed its page layout. Fixed, and the app now reports an error instead of silently showing only Hugging Face results.
• Hugging Face repos that ship the same model both whole and split no longer show double the size.

v0.5.0 — First stable release

Ollama Manager has grown from a simple Ollama service toggle into a full control panel for
a local AI stack, and after several release-candidate cycles (v0.4.x) and daily use without
issues, this is the first release marked stable.

Highlights

• Multi-host model management — switch between localhost and any LAN host (e.g. a BC-250
  box) from a dropdown, no config-file editing. Service control (start/stop/autostart) always
  stays local, since systemd can't be driven remotely.
• Model aggregator (LiteLLM) — one OpenAI-compatible endpoint that routes to models across
  all configured hosts, installed and managed with one click (no Docker). Includes a
  Continue.dev config generator so VS Code picks up every host's models automatically.
• Open WebUI integration — install, start/stop and autostart a full browser chat UI as a
  systemd --user service, with an offer to install the optional system packages
  (ffmpeg/pandoc/zstd) needed for voice transcription and document RAG.
• Advanced tab — direct control over 9 Ollama environment variables (OLLAMA_KEEP_ALIVE,
  OLLAMA_CONTEXT_LENGTH, OLLAMA_NUM_PARALLEL, Vulkan backend, iGPU toggle, and more) without
  hand-editing systemd override files.
• Seven languages — English (default), Polish, German, Spanish, French, Italian, Portuguese,
  switchable live from the window, no restart.
• Two install paths — a no-root install.sh (installs to ~/.local/share) or a proper
  .deb package (build-deb.sh) with apt-managed dependencies.
• NEW in this release: update checks. One-click "Check for updates" for both Ollama
  (compares against the latest GitHub release) and Open WebUI (compares against the latest
  PyPI release) — each offers to install the update immediately if you're behind. Useful since
  newer models sometimes require a newer Ollama to even load.

Upgrading

​ git pull ./install.sh # or: ./build-deb.sh && sudo apt install ./ollama-manager_*_all.deb ​

Requirements

PyQt6 + requests, systemd + polkit (pkexec), curl — see the README for the
full list and optional dependencies.

v0.4.21 (RC4)

What's new

• English is now the default UI language instead of Polish. New users
  (fresh install, no saved language preference) now see the app in English
  by default — the app is meant to be usable by anyone, and English is
  understood more broadly. You can still switch to any of the 7 supported
  languages from the server bar at any time.
• Open WebUI install now offers optional extras. Right after a
  successful Open WebUI installation, the app asks whether to also install
  ffmpeg, pandoc, and zstd via apt/pkexec — these unlock audio
  transcription and PDF/Word document import into RAG. Previously this was
  only documented as a manual step in the README; WebUI works fine without
  them either way.

Status

This is the last planned release candidate before the stable 0.5.0
release. From here on, work shifts to bug-hunting and stabilization —
no new features planned before stable.

Ollama Manager v0.4.12 (RC3)

Changes since v0.4.5 (RC2).

Advanced tab: redesigned

• Compact form (name + field/checkbox, no per-row descriptions) with a single "Save" button that applies every field at once and restarts the service once - instead of a separate save+restart per field.
• New "Apply recommended values" button pre-fills the form with a suggested BC-250-oriented profile (context 16k, Vulkan+iGPU on, q8_0 KV cache, etc.) for review before saving - it never writes anything by itself.
• New free-form custom parameters field for any environment variable not covered by the built-in list - anything already in override.conf that doesn't match a known field shows up there instead of being lost.
• Full variable descriptions moved to a new separate "Help" tab, keeping Advanced itself short.
• Two new variables: GGML_VK_VISIBLE_DEVICES (Vulkan device index - needed on some APUs) and OLLAMA_GPU_OVERHEAD (VRAM reserved for the rest of the system - picker with "desktop" and "server" presets, desktop selected by default since this is a GUI app).

Fixed: override.conf parser corrupting manually-edited entries

If override.conf was ever hand-edited (e.g. via systemctl edit ollama) using multiple variables on one Environment= line, or single quotes, the old parser silently mangled it - variables got merged together or picked up a stray quote in their name, so toggles like Vulkan stopped reading/writing correctly. The parser now uses shlex to parse each line the same way systemd itself does.

Model download: size, quantization, and a memory estimate

• New quantization picker (default/Q8_0/full F16) and size picker when pulling a model - the size list is now specific to the selected model, verified against ollama.com/library (no more suggesting sizes a model doesn't actually have, e.g. Ornith only ships 9b/35b, not the old generic 1b-70b list).
• Live, labeled-as-approximate memory estimate (weights + KV cache + buffers, then a recommended-RAM figure) as you pick a size/quantization - assumes q8_0 KV cache with a warning that this needs OLLAMA_FLASH_ATTENTION=1 to actually apply.
• Suggested model list gained GPT-OSS, DeepSeek-Coder-V2, and Ornith.
• Dropdowns widened so all suggestions show without scrolling.

Installed models: now a table

The "Zainstalowane modele" list shows disk size and an estimated/live VRAM column per model (real usage if currently loaded, otherwise the same calculator using the model's real quantization metadata from Ollama's API).

Installer

• install.sh now offers to install missing curl or python3-pip via apt (with confirmation) instead of just printing instructions and exiting.
• build-deb.sh adds curl to the package's Depends:.

Other

• Wider minimum window size (fits the taller Advanced/download content).
• README got a screenshot and a "still beta, actively developed" notice.

Requirements

Unchanged - Python 3, PyQt6, requests, systemd, polkit. See the README.

This is a release candidate

Same caveat as before - please open an issue if something breaks.