AI Settings¶
The AI Settings dialog is where you configure the providers and models that power MAGDA's AI features: the AI Assistant chat, the per-device sound-design panels, the Media Library's Sample Analyzer, and Stem Separation.
Open it from Settings > AI Settings. The dialog has three tabs.
| Tab | Purpose |
|---|---|
| Cloud | Register cloud API providers (OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter) |
| Config | Choose which providers each agent uses, and how they trade off quality, speed, and cost |
| Models | Download and manage every optional local model, grouped by category |
Cloud¶
The Cloud tab registers API keys for hosted LLM providers.
To add a provider:
- Pick the Provider from the dropdown (OpenAI, Anthropic, Gemini, DeepSeek, or OpenRouter).
- Enter your API Key.
- Click Test to verify the key, then Add to register it.
Registered providers appear in the list below. You can register several and switch between them from the Config tab. Click Remove to delete a provider. A provider already in the list is greyed out in the dropdown so you do not add it twice.
Keys stay on your machine
API keys are stored locally in your MAGDA configuration and used only to call the provider you entered them for.
Config¶
The Config tab decides which providers the AI agents actually use. The Setup dropdown at the top picks how much control you want:
- Simple — one decision that applies to every agent.
- Advanced — a provider and model per agent role.
Simple setup¶
| Setting | Description |
|---|---|
| Mode | Local (every agent uses a local model), Cloud (every agent uses a cloud provider), or Hybrid (a mix) |
| Source | In Local mode: Embedded (the built-in llama.cpp engine, configured under Models > Local LLM) or Local server (an OpenAI-compatible server you run yourself) |
| Provider | Which registered cloud provider to use, shown for Cloud and Hybrid modes |
| Optimize | Bias agent selection toward Quality, Speed, or Cost |
In Local mode the provider and optimize options are hidden, since everything runs locally.
Local Server (OpenAI-compatible)¶
With Source set to Local server, MAGDA talks to any server that speaks the OpenAI API — LM Studio, Ollama, GPUStack, or a plain llama.cpp server. The default address is Ollama's (http://localhost:11434/v1); MAGDA normalises whatever form you give it, with or without the /v1 suffix.
- Model — pick the model the server should use. Refresh queries the server's model list; you can also type a model id directly.
- An API key/token is only needed for servers that require one (for example GPUStack); the others run without it.
One server and one model serve all agent roles, the same way the embedded model does.
Advanced setup¶
Advanced replaces the single mode with a grid, one row per agent role. Each row has its own provider and model, drawn from whatever you registered on the Cloud tab and whatever local backends are available. Switching from Simple to Advanced seeds the grid from your simple settings, so you can start there and adjust only the rows you care about.
| Agent | Handles |
|---|---|
| Command | Turning chat requests into DSL operations on tracks, clips, and devices |
| Music | Generating musical content — notes, patterns, parts |
| Faust | Writing Faust DSP code for custom effects |
| Chord | Chord progressions and harmony |
| Controller | Controller profiles and scripts |
| Theme | Colour themes |
Choosing a local provider gives a model dropdown of Embedded or Server. The Command row has a third option, Fast Inference (Command) — see below.
Fast Inference (Command)¶
Fast Inference runs Command requests entirely on your machine with a purpose-built model rather than a general-purpose LLM. It is not a chat model: it reads one command and emits the DSL directly, with no streaming and effectively no wait.
Because it is a single-command tagger, it deliberately ignores conversation history — each request is read on its own. For follow-ups that depend on what you said earlier ("now make it louder"), a general model is the better fit.
Fast Inference works out of the box using a small built-in model. Installing the larger Command model from the Models tab replaces it and markedly improves how well natural phrasing is understood. Cloud providers never use either one.
MCP Tools¶
Faust DSP validates AI-generated Faust code before MAGDA loads it. MAGDA starts the MCP server on demand through npx, so there is nothing to install by hand, but npx must be on your system PATH (it ships with Node.js). See Effects — Faust validation for details.
Models¶
Every optional download lives on the Models tab, behind a Category dropdown. Categories appear only when your build supports them — the ONNX-backed ones (Sample analysis, Stem separation, Command model) are absent from the Intel macOS build.
Downloads and removals take effect straight away; they do not wait for OK.
Local LLM¶
Configures the embedded llama.cpp engine for fully offline AI, with no network calls and no API key.
- Download Model — downloads the MAGDA model (fine-tuned for DAW operations) from HuggingFace with a progress indicator.
- Model (.gguf) — the path to the model file. Use the browse button to point at any GGUF model of your own.
- GPU Layers — Auto (GPU) offloads everything the platform can take (Metal on macOS, CUDA on supported builds), CPU Only runs without the GPU, and Custom lets you enter a layer count for cards that run out of memory on Auto.
- Context — the context-window size in tokens (default
4096). - Load Model / Unload — load the model into memory or unload it to free resources.
- Load model on startup — when enabled, the model loads automatically each time MAGDA launches.
Sample analysis¶
The Sample Analyzer is the audio-tagging model behind the Media Library's semantic search and "find similar sounds". It is an optional download, separate from the LLM models.
- The page shows the current install state and the download size.
- Download Sample Analyzer — fetches the model. Without it, the Library still supports filename, tag, family, shape, key, and BPM filtering, but not text-based semantic search.
- Load / Unload — bring the analyzer into memory or release it. It also preloads in the background the first time you enter Library mode.
For how the analyzer is used once installed, see Media Library > AI Sample Analyzer.
Stem separation¶
Weights for the machine-learning engines behind Split into Stems. The HPSS engine needs nothing here — it is pure signal processing and always available.
| Model | Size | Splits into |
|---|---|---|
| Demucs (4-stem) | 301 MB | Drums, Bass, Other, Vocals |
| Spleeter (2-stem) | 74 MB | Vocals, Accompaniment |
Each row shows its install state, a progress bar while downloading, and a Download / Remove button, alongside a link to the HuggingFace repository the weights come from. Models location shows where they are stored.
Choosing an uninstalled engine from a clip's Split into Stems menu opens this page directly.
Command model¶
The larger model behind Fast Inference: a DeBERTa-v3 encoder fine-tuned to read a command and tag its intent and parameters.
- It is not bundled with MAGDA — the download is 429 MB.
- Without it, Fast Inference still works using the small built-in model. With it, natural phrasing is understood far more reliably.
- Models location defaults to MAGDA's data folder. Use Browse... to keep the weights on another drive, or Reset to go back to the default.
- The Download button turns into Cancel while a download is running, and Remove once installed. The location cannot be changed mid-download.
- A link to the source HuggingFace repository sits under the model name.