Ollama Manager

A small Flask web app that makes adding, removing and tuning Ollama models painless.
No more hand-writing Modelfiles.

---

What it does

• Add a .gguf model by filling in a form. It builds the Modelfile for you, shows a live preview, and adds the model to Ollama behind a progress modal.
• Prompt formats built in: auto-detect from the file name, or choose ChatML, Llama 3, Mistral, Gemma, Phi-3, Alpaca, or the template embedded in the GGUF.
• Browse the .gguf files in your models folder.
• Installed models: list, view the Modelfile, copy the name, and remove (with a confirmation dialog).
• Daemon controls: start, stop and restart the Ollama service, with a live VRAM progress bar.
• Timeout and memory: change how long idle models stay loaded, see what is in memory, and unload a model on demand.
• Settings: flash attention, KV cache type, parallel requests and max loaded models.
• Cheat sheet of useful Ollama commands with copy buttons, and an activity log of every model added or removed.
• Ten colour themes, large buttons and styled tooltips, built on Bootstrap 5 and Bootstrap Icons.

Requirements

• Linux with systemd (developed on Lubuntu 22.04)
• Ollama installed as a systemd service named ollama
• Python 3.10+ and Flask (pip install flask)
• zenity, only if you use run.sh (for the password prompt)
• An NVIDIA GPU with nvidia-smi, only for the VRAM bar (everything else works without it)

Windows and macOS are not supported. The daemon controls rely on systemd and the launcher on sudo and zenity, so a port would need a substantial rework.

Quick start

git clone <your-repo-url> ollama-manager
cd ollama-manager
./run.sh

Then open the address it prints, normally http://127.0.0.1:5782/. If port 5782 is busy, the app tries the next free port.

Two ways to launch

Command	Daemon start/stop/restart and settings	Everything else
./run.sh	Works. Asks for your sudo password in a dialog, then runs the app as root.	Works
python3 app.py	Disabled, because the app needs root to control a system service.	Works

When launched with run.sh, files the app creates in logs/ and data/ are handed back to your user when it exits.

Configuration

The constants at the top of app.py are meant to be tweaked:

Constant	Purpose
DEBUG_MODE	False keeps the console quiet. True prints request logs and writes extra audit lines to the log file.
PORT, PORT_TRIES	Starting port and how many following ports to try.
DEFAULT_MODELS_DIR	Where the Browse button looks for .gguf files.
DEFAULT_TEMPERATURE, DEFAULT_CONTEXT, DEFAULT_SYSTEM_PROMPT	Form defaults.
LOG_MAX_BYTES	The log is trimmed to its newest half once it exceeds this size (100 KB by default).

Form values are remembered in your browser's local storage.

Project layout

app.py              entry point and constants
run.sh              launcher with a zenity sudo prompt
assets/
  html/             the single page
  css/              Bootstrap plus modular, token-based CSS and themes
  js/               page logic
  py/               helper modules (functions_*.py) and the root helper
  img/              README image
logs/               ollama_manager.log (created on demand, git-ignored)
data/modelfiles/    generated Modelfiles (created on demand, git-ignored)

Security notes

• The server binds to 127.0.0.1 only and rejects requests whose Host header is not localhost.
• Every POST must be JSON, which blocks simple cross-site form posts.
• Model names, paths and settings are validated, and no shell is ever used to run commands.
• Root actions go through a small helper that accepts only start, stop, restart and a whitelist of five Ollama settings.
• Running the app as root is a deliberate convenience. Keep it on a machine and network you trust.

Notes

• Adding a model only runs ollama create. It never loads or runs a model.
• Changing a daemon setting restarts Ollama, which unloads any loaded model. The app asks you to confirm first.
• Settings are written to a systemd drop-in, /etc/systemd/system/ollama.service.d/zz-manager.conf.