🦙 LLamaStudio

A desktop chat interface and local server manager for llama.cpp, crafted with FastAPI + HTMX for ultra-lightweight, zero-framework execution.

LLamaStudio is a self-contained local workspace that manages model lifecycles, features a smart VRAM estimator, scans local folders, and lets you search and download models directly from the Hugging Face Hub.

---

📸 Screenshots & Showcase

Terminal tour

Watch the MP4: lls tui, then Discover, Models, a live chat, and inference logs—all using an already loaded model. Regenerate the recording after changing the palette.

1. Main Chat Dashboard

A Pop!_OS-harmonized dark interface with streaming, collapsible markdown reasoning (thinking) processes, and real-time agentic tool execution logs.

2. GGUF Model Browser & Settings

A dynamic local model explorer that scans your directories and lets you adjust context length, GPU offload layers, CPU threads, flash attention, and KV cache quantizations on the fly.

3. Hugging Face Discover Hub

Browse the entire Hugging Face GGUF catalog. Features a Smart VRAM Offload Estimator calibrated to your hardware, and a floating background download progress card with live speed (MB/s), ETA, and cancel controls.

---

✨ Key Features

• ⚡ Zero Node Modules: Built with HTMX, Tailwind CSS (via CDN), and Vanilla JS. It is incredibly fast, responsive, and has a memory footprint of just a few megabytes.
• 🧭 Hugging Face Discover Tab: Search the public Hugging Face Hub for GGUF models directly inside the app, view readmes, select quantizations, and download files in the background.
• 🚀 Smart VRAM Estimator: calulated specifically for your hardware (fits fully on RTX 5090 32GB VRAM, partial offload warning, or heavy CPU fallback warning).
• 📂 Automatic Model Scanning: Scans standard directories (like ~/.lmstudio/models) automatically on startup or via a one-click rescan button.
• 🪐 Process Lifecycle Manager: The underlying llama-server process only spins up when you explicitly load a model, releasing all system resources and GPU VRAM instantly when you click "Eject".
• 🔧 Configurable Workspace Sandboxing: Supports sandboxed agentic tool use (file read/write, commands, etc.) with real-time logs in the UI. Workspace and permission defaults are stored in the first-class app config.
• 👁️ Multimodal Media Chat: Drag raster images or WAV/MP3/FLAC audio into chat, or attach workspace media with lls oneshot --image / --audio; FLAC is normalized to llama.cpp-compatible WAV with ffmpeg.
• 🎙️ Local Push-to-Talk: Click the microphone once to record and again to stop. A managed whisper.cpp service transcribes locally and places editable text in the chat box; audio is never sent to the chat model.
• 🖥️ XDG-Compliant Persistence: App config, conversations, and first-class model profiles are stored outside the codebase directory in standard ~/.config/llamastudio/ with automated backward-compatible migrations.
• 📦 Full Linux & macOS Portability: Server binaries and model directories are resolved dynamically on startup.

---

🛠️ Installation & Setup

LLamaStudio is compatible with Linux and macOS out-of-the-box. Choose your OS and python virtual environment preference below.

🐧 1. Linux Installation

Prerequisites

1. Python 3.10+ (Recommended: Python 3.13)
2. llama.cpp built from source (or pre-compiled binary):
   • By default, the app dynamically looks for the llama-server binary globally on your system PATH or locally inside your home directory at ~/llama.cpp/build/bin/llama-server.

Environment Setup

Option A: Install from PyPI

pip install llamastudio

Option B: Conda / Miniconda from source

# 1. Clone the repository
git clone https://github.com/gnulnx/LlamaStudio.git
cd LlamaStudio

# 2. Create and activate a conda environment
conda create -n llamastudio python=3.13 -y
conda activate llamastudio

# 3. Install LlamaStudio and its dependencies
pip install -e .

Option C: Python Virtualenv (venv) from source

# 1. Clone the repository
git clone https://github.com/gnulnx/LlamaStudio.git
cd LlamaStudio

# 2. Create and activate a python venv environment
python3 -m venv .venv
source .venv/bin/activate

# 3. Install LlamaStudio and its dependencies
pip install -e .

🖥️ Linux Desktop Launcher Integration (Optional)

To integrate LLamaStudio directly into your Linux Application launcher menu (e.g., GNOME / Pop!_OS):

# 1. Copy the desktop file to your local applications directory
cp llamastudio.desktop ~/.local/share/applications/

# 2. Copy the custom SVG icon to your local icons directory
mkdir -p ~/.local/share/icons/hicolor/128x128/apps/
cp llamastudio.svg ~/.local/share/icons/hicolor/128x128/apps/

# 3. Update your desktop database and icon cache
update-desktop-database ~/.local/share/applications/
gtk-update-icon-cache -f -t ~/.local/share/icons

Note: If you are using a virtualenv, edit the executable path inside ~/.local/share/applications/llamastudio.desktop to point to your specific .venv/bin/python interpreter.

---

🍏 2. macOS Installation

Prerequisites

1. Python 3.10+
2. llama.cpp installed globally via Homebrew (highly recommended for macOS):

   brew install llama.cpp
   

   (This automatically places the llama-server binary globally on your system PATH, which LLamaStudio will auto-detect immediately!)

Environment Setup

Option A: Install from PyPI

pip install llamastudio

Option B: Conda / Miniconda from source

# 1. Clone the repository
git clone https://github.com/gnulnx/LlamaStudio.git
cd LlamaStudio

# 2. Create and activate environment
conda create -n llamastudio python=3.13 -y
conda activate llamastudio

# 3. Install LlamaStudio and its dependencies
pip install -e .

Option C: Python Virtualenv (venv) from source

# 1. Clone the repository
git clone https://github.com/gnulnx/LlamaStudio.git
cd LlamaStudio

# 2. Create and activate venv
python3 -m venv .venv
source .venv/bin/activate

# 3. Install LlamaStudio and its dependencies
pip install -e .

---

🪟 3. Windows Installation

Note: Native Windows execution is currently untested. However, you can run LLamaStudio on Windows seamlessly via WSL2 (Windows Subsystem for Linux) by following the standard Linux Installation guide above.

Pull requests extending native Windows support (e.g., resolving .exe binaries) are highly welcome!

---

🚀 Running the Application

Option A: Via Unified CLI (lls - Recommended)

You can link and install LlamaStudio's CLI utility locally to control the desktop app and server seamlessly:

# Start the desktop application server and open browser UI
lls start

Option B: Via App Launcher Command

After installing from PyPI or source, run:

llamastudio

Via Application Menu (Linux)

Search for LLamaStudio in your desktop search bar (press Super, type "Llama") and click to launch!

---

🛠️ Unified Command-Line Interface (lls)

LlamaStudio features a CLI built using rich-click for visual dashboards and operational efficiency.

CLI Subcommands Reference

Command	Usage	Description
start	lls start	Starts the desktop app and opens the browser to the right first-run/chat/models/discover view.
tui	lls tui [--view discover|models|chat|logs]	Interactive terminal interface with mouse, keyboard, and adaptive layouts. Starts the backend without opening a browser when needed.
reload	lls reload	Gracefully restarts the desktop FastAPI application backend.
status	lls status	Visual dashboard of FastAPI backend status, loaded model parameters, and GPU memory (VRAM).
ls	lls ls	Prints an elegant table of all GGUF models scanned across local directories.
load	lls load [MODEL]	Boots the server with a GGUF model. If MODEL is omitted, prompts you with an interactive menu.
eject	lls eject	Gracefully unloads the active model to free GPU and CPU RAM.
oneshot	lls oneshot [--image PATH] [--audio PATH] [--no-thinking] [--max-tokens N] "prompt"	Streams text, optional reasoning, tool calls, and multimodal workspace images/audio directly in your terminal. Use --no-thinking for low-latency direct answers.
speech status	lls speech status	Shows the local Whisper installation, selected model, compute mode, and server state.
speech install	lls speech install [--model small.en]	Installs pinned, checksum-verified whisper.cpp Linux binaries and a local Whisper model.
speech load/eject	lls speech load [MODEL] [--gpu|--cpu]	Starts or stops the persistent speech-to-text server independently of the chat model.
speech transcribe	lls speech transcribe AUDIO	Transcribes a workspace audio file locally, with optional language and English translation controls.
speech record	lls speech record [--device default]	Starts terminal microphone capture immediately; press Enter to stop and print the transcript.

Set up push-to-talk once, then use the microphone beside the chat input:

lls speech install --model small.en
lls speech status
lls speech record

The browser control is a toggle, not a hold action. The first click starts recording, the red stop button ends it, and the transcript is inserted without auto-sending so it can be corrected first. Browser microphone access requires the loopback URL (http://127.0.0.1:8765) or HTTPS.

For low-latency vision classification, disable reasoning and keep the answer budget small:

lls oneshot --no-thinking --temperature 0 --max-tokens 32 \
  --image camera-frame.png "Answer in 10 words or fewer: what is ahead?"

For audio transcription or translation with an audio-capable model and projector:

lls oneshot --no-thinking --audio recording.flac \
  "Transcribe this audio, translate it to English, and respond briefly."

For example, to boot a model interactively:

$ lls load
Available Scanned Models:
  1. Qwen3.6-35B-A3B-UD-Q5_K_M (25.2 GB)
  2. gemma-4-26B-A4B-it-Q8_0 (25.0 GB)
  3. DeepSeek-R1-Distill-Qwen-32B-Q5_K_M (21.7 GB)

Select a model number to load: 3
Loading model 'DeepSeek-R1-Distill-Qwen-32B-Q5_K_M'...

---

Terminal interface

lls tui
lls tui --view chat
lls tui --no-start     # Connect only; fail if the backend is offline

The TUI shares the web app's backend, model profiles, downloads, and saved conversations. Closing it leaves the backend and loaded model running. Discover supports Hugging Face search, sorting, README/details/files, GGUF selection, download progress, and cancellation. Models provides filtering, rescan, saved load settings, load/eject, and confirmed deletion. Chat supports multiline text, saved conversations, streamed Markdown, collapsible reasoning, and tool activity. Logs tails the application or inference server, with filtering and follow control.

The header adds a two-line wordmark, the primary GPU, live device-wide VRAM usage, inference state (including CPU/GPU mode), and the active model. NVIDIA memory usage refreshes every three seconds; unsupported telemetry is shown as unavailable, not zero. Apple unified memory is labeled separately. Narrow or short terminals use a two-line status summary instead of the full header.

Use F2–F5 to switch sections, Tab/Shift+Tab to move between controls, arrows and Enter to select, and F1 for help. In chat, Enter inserts a newline and Ctrl+S sends. Ctrl+R refreshes; Ctrl+Q quits. Every workflow also has mouse controls. In terminal multiplexers, forward shortcuts to the application (for example, use zellij's locked mode if its bindings intercept Ctrl+S). On macOS you may need Fn with function keys, or use the navigation buttons.

Layouts adapt from a navigation rail and side-by-side panels to a navigation bar and separate list/detail views below 110 columns. An 80×24 terminal is supported; larger windows show more columns and model information. No Nerd Font, terminal image extension, or graphics protocol is needed. Standard terminal colors are used when true color is unavailable, and NO_COLOR is respected. For SSH, run lls tui on the host running LlamaStudio, using ssh -t when launching directly. The backend's filesystem and GPU are the ones shown in the TUI.

Colors look wrong over SSH? SSH may pass TERM=xterm-256color without COLORTERM=truecolor, causing the TUI to use a reduced 256-color palette. Dark surfaces can collapse to black and accents can shift noticeably. If your local terminal supports true color (for example, iTerm2), launch the TUI on the remote host with:

COLORTERM=truecolor lls tui

This also works with --theme default, --theme slate, or another theme choice. Set the variable in the remote shell where lls runs; setting it only in your local shell does not ensure SSH forwards it. Use this override only with a true-color-capable terminal. Native sessions may look correct while SSH sessions from the same terminal look different.

Choose an appearance at launch, or press F6 to open the Appearance chooser:

lls tui --theme default  # Original LlamaStudio purple (the default)
lls tui --theme light    # Pale surfaces with purple accents
lls tui --theme dark     # Charcoal surfaces with neutral accents
lls tui --theme slate    # Named preset: slate surfaces with blue accents
lls tui --theme system   # System adapter, native light/dark preference; Default if unsupported

Selections in the chooser apply to the current TUI session; they are not saved as a desktop-wide preference. Ctrl+P reloads the selected source. Switching or reloading preserves your current section, selection, and unfinished chat draft. --theme also works with --screenshot.

System follows the native light/dark preference, using the bundled Light or Dark palette. On GNOME-family desktops (including GNOME-based Pop!_OS), it reads org.gnome.desktop.interface color-scheme through gsettings. An explicit prefer-light or prefer-dark wins; default or a missing legacy key falls back to gtk-theme (for example, Pop-dark). Theme names containing a separate dark component select Dark; other names select Light. This is a naming convention, not an attempt to parse arbitrary GTK stylesheet colors.

On macOS, System reads the global AppleInterfaceStyle preference through defaults; an absent key means Light. Both adapters reread the effective preference every two seconds while System is selected, including changes made by automatic appearance scheduling. Reads run off the UI thread, have bounded timeouts, and never modify desktop settings or request automation permission. Unsupported/headless environments display a notice and use Default without a polling timer. Over SSH, preferences belong to the host running the TUI.

Slate is the first named preset, preserving the original slate-blue Dark design.

On Omarchy, System uses the active desktop theme's full palette instead of bundled Light or Dark, detected ahead of the native adapters. It activates only when ~/.local/state/omarchy/current/theme/colors.toml exists and omarchy-theme-color is on PATH, and reads colors only through that command with the same bounded timeout. Status colors (success, teal, warning, error) keep their Default values, and the theme's mode selects light or dark. The adapter checks the theme file every quarter second, reruns the command only after Omarchy replaces that file, and keeps the current theme while a switch is in progress.

Default colors live in app/tui/palette.json, shared by the stylesheet, header, capability badges, model status, and logs. Edit that file and press Ctrl+P in the TUI to reload it without restarting, refetching model lists, or losing a chat draft. Purple marks actions/selections, green and teal mark status/compatibility, gold marks capabilities/tool activity/warnings, and lilac grays mark secondary text. Layout rules stay in app/tui/studio.tcss.

For a separate experiment (or an installed wheel), create a JSON file inside the configured workspace containing just the roles you want to override, for example:

{
  "primary": "#8b5cf6",
  "focus": "#9d6bff",
  "border": "#6d3df0"
}

Run lls tui --palette colors.json; Ctrl+P reloads that file. Colors must be six-digit #RRGGBB values. Invalid edits keep the last working palette and show an error. --palette also works with --screenshot. Terminal color capability and NO_COLOR still apply; a palette cannot add true color to a terminal that does not support it.

Custom colors inherit the selected bundled theme and its light/dark mode: lls tui --theme light --palette colors.json. An explicit palette wins over System, bypasses system detection, and uses Default as its base. F6 also offers Custom when launched with --palette, so you can compare it against the bundled themes and return to your file.

Theme adapters are defined in app/tui/themes.py. Each resolves a validated ResolvedTheme containing palette, dark-mode flag, source, and an optional user-facing notice. Bundled and custom themes use this same contract. To add a system integration, register an ordered detection factory in SYSTEM_ADAPTERS; it returns an adapter when supported, otherwise None. Detection runs only when System is selected. An adapter may request polling with refresh_interval; None means no background work. The adapter owns bounded I/O and revision caching, and raises OSError or ValueError on an unreadable source. It must not touch widgets, start its own timer, or mutate application settings. Use the workspace path validator for user-provided palette paths; any platform integration's fixed system-file access must be explicitly scoped and reviewed.

The shell serializes live resolution off the UI thread, ignores obsolete results, and applies CSS and inline colors together. Initial system-source failure falls back to Default; live failure retains the last good theme, reports the error once, and retries on subsequent polls. Leaving System stops its timer. Static sources reload with Ctrl+P. This keeps platform-specific settings out of the CLI and views.

This first version handles text chat; media input remains in the web app. Split GGUF shards are identified but not offered as individual model downloads: fetch the complete set from the linked Hub repository. The Discover memory bar compares GGUF weight bytes with reported GPU/unified memory, not guaranteed load capacity; context cache and runtime allocations need additional memory. The backend still shares one active conversation, so avoid sending simultaneously from the web app and TUI.

Capture the real interface for visual review without a terminal:

lls tui --screenshot .runtime/tui/discover.svg --size 190x52
lls tui --view models --screenshot .runtime/tui/models.svg --size 80x24

Screenshots must be new .svg files within the configured workspace. They contain the live backend's data; use an appropriate conversation before sharing chat captures.

Recording the demo

From the checkout, with the backend running and a model already loaded:

lls status
lls demo-tui
# Optional palette experiment and separate output:
lls demo-tui --palette colors.json --output imgs/tui/experiment.mp4

The reproducible script is app/tui/demo.tape; edit its pauses, typing, and navigation to adjust the tour. The command uses the current TUI and palette—not canned screens—and produces imgs/tui/demo.mp4, a still imgs/tui/demo-poster.png, and imgs/tui/demo.gif. The lightweight README animation links to the full-quality MP4.

Install VHS (tested with 0.11.0), ttyd, FFmpeg (including ffprobe), Bash, and Chrome/Chromium. The tape uses DejaVu Sans Mono. No recording dependencies are added to the app's runtime requirements. Allow a few minutes for capture, encoding, and validation.

The tour opens an edge-to-edge terminal matching the TUI palette's background, with three window dots and no outer frame. It types lls tui, browses all four sections, and requests a short real reply. It waits for Hub results and chat completion; network speed and the loaded model affect its duration and content. It never loads, ejects, downloads, or changes a model profile. A temporary demo conversation is created and deleted afterward; the previously active conversation is restored unless you switched it elsewhere. Keep other clients idle during recording.

Review before publishing: real conversation titles, model names, filesystem paths, and logs can appear in the video. This is a live capture, not an anonymizer. The recorder enables true color only in its child environment, preserves shell history, and exports 1920×1080 H.264/yuv420p with fast-start playback. The GIF is a 960×540, 3 fps, 96-color looping preview, kept below 10,000,000 bytes for GitHub's image limit. All outputs are staged; the MP4 passes metadata and full-decode checks and the GIF passes a full-decode check before replacing the previous recording. Changing the palette and rerunning the same command refreshes the demo without hand-editing the video.

⚙️ Configuration & Customization

The application runs fully out-of-the-box with no manual configuration. On first launch, LlamaStudio creates its runtime config under:

~/.config/llamastudio/
  config.json
  model_profiles.json
  conversations.json
  logs/

config.json stores app defaults, model search directories, workspace permissions, and launch state. model_profiles.json stores first-class per-model load and inference profiles. Older model_settings.json files are migrated automatically.

---

🛡️ Workspace Sandboxing & Embodiment

By default, LlamaStudio restricts agent tools (like reading, writing, and listing files) to the configured workspace directory to prevent accidental path traversals. For CLI launches, the first-run workspace defaults to the directory where lls start was run.

Workspace configuration is saved in ~/.config/llamastudio/config.json. Environment variables are still supported for advanced/bootstrap overrides, but normal users should not need a .env file.

Developer details for the config/profile architecture live in DEV.md.

---

🧪 Testing Suite

LlamaStudio features both standard unit tests and comprehensive GGUF integration tests.

1. Standard Unit Tests

Verify local installation and confirm backend routing, regex parsing, and sandboxing safety behaviors by running our mock-based test suite:

python -m unittest discover tests

2. GGUF Model Integration Tests

For local environments containing active GPUs and downloaded models, you can run the full multi-model GGUF tool-calling integration suite to verify real-time execution robustness across various chat templates:

# Run GGUF model integration tests locally
./tests/test_all.sh

(These tests are automatically skipped in standard CI/CD environments and default pytest runs using @pytest.mark.skipif to keep pipeline checks fast.)

3. Terminal UI tests

The default suite includes HTTP/SSE contract tests, CLI lifecycle tests, and Textual Pilot tests exercising actual widgets with isolated backend responses:

python -m pytest tests/test_tui.py tests/test_tui_client.py tests/test_tui_cli.py -q

For a real model response, reasoning, tool activity, persistence, and screenshots:

lls status  # A model must already be loaded
RUN_TUI_LIVE=1 python -m pytest tests/test_tui_live.py -q -s

This opt-in test creates and removes only its own conversation, restores the previous selection, and writes review SVGs under .runtime/tui/.

---

🏗️ Project Structure

LlamaStudio/
├── pyproject.toml         # Package metadata, CLI entrypoint, and dependencies
├── DEV.md                 # Development notes for runtime config and profiles
├── llamastudio.desktop    # GNOME/Linux desktop launcher metadata
├── llamastudio.svg        # Custom application vector icon
├── app/
│   ├── config.py          # Settings & dynamic path configurations
│   ├── config_store.py    # First-class runtime config and model profiles
│   ├── main.py            # FastAPI backend endpoints & routing
│   ├── chat.py            # Conversations registry, templates & chat streaming
│   ├── downloader.py      # Async background download manager (chunked writes)
│   ├── model_manager.py   # Scans local paths and Hugging Face Hub
│   ├── server_manager.py  # llama-server subprocess process lifecycle controller
│   ├── logger.py          # Centralized logger
│   ├── tools.py           # Sandboxed local workspace tools for LLM agent use
│   └── templates/
│       └── index.html     # Interactive HTMX frontend interface
├── tests/
│   └── *.py               # Unit and integration-adjacent test coverage
└── imgs/
    ├── chat_interface.png # Screenshot: Main Chat interface
    ├── model_settings.png # Screenshot: Model explorer & settings
    └── discover_models.png# Screenshot: HF Discover & Downloader panel

---

📄 License

LLamaStudio is open-source software licensed under the MIT License.