MAGDA - Free Open Source DAW with Integrated AI (v0.9.0)

[ConceptualMachines]

0.5.0 is out — automation lands.

What's new:

• Automation — every parameter MAGDA knows about (track volume, pan, sends, plugin parameters) can now be drawn, recorded in real time, or written by the AI agent from a plain-English prompt
• QWERTY MIDI keyboard — toggle in the transport, your computer keyboard plays the selected instrument track. Piano-style layout, Z/X to shift octave
• Simplified Chinese — first non-English locale, full localization pipeline via Crowdin. If you speak a language that isn't there yet, you can join the project and contribute
• Configure Parameters — right-click a plugin in the browser to inspect every parameter as a row. Heuristic infers units/ranges automatically; AI Detect handles the ambiguous ones; manual overrides for everything else
• Click-to-detect (Learn) — for plugins with many parameter pages, click Learn on a device slot then touch any control in the plugin's window — MAGDA jumps straight to its slot
• Smaller wins: multi-sample drag-and-drop, Duplicate Track (Content Only) ⌘⌥D, AI agent now respects track selection, Check for Updates from Help menu, plus assorted fixes

Video walkthrough:
Download: https://magda.land

Touch/Latch automation modes are coming next — current pass is Write only. If you try the new automation flow and have feedback, let me know.

[pekbro]

Awesome thanks!

*Midi sends between plugins do indeed work well on windows.

[satYatunes]

I tried the last version. The local AI kept thinking for a very long time. Is that expected? I got bored and shut it down, uninstalled. I will wait for some time before re-installing. AI is not the only thing that I am interested in btw. Sample auditioning syncing to project tempo and key is not there. Having said that I totally understand it's at 0.5.0 and evolving very fast. I have it on my radar

[ConceptualMachines]

I tried the last version. The local AI kept thinking for a very long time. Is that expected? I got bored and shut it down, uninstalled. I will wait for some time before re-installing. AI is not the only thing that I am interested in btw. Sample auditioning syncing to project tempo and key is not there. Having said that I totally understand it's at 0.5.0 and evolving very fast. I have it on my radar

Thanks for trying it. The local model we ship is already a 7B, about as small as is usable, so the speed really comes down to the machine. Out of curiosity, what's your setup? Local 7B inference really wants a GPU; on a plain CPU it'll feel like it's hung even when it's working. If you want responsive AI without that constraint, a hosted API key via BYOAK is the realistic path.
Sample auditioning synced to tempo and key is a nice suggestion — noted.

[satYatunes]

Code: Select all

OS           Microsoft Windows 11 Pro (Build 26200)
CPU          Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz
RAM          32GB (16GB + 16GB) @ 2400MHz
GPU          Intel(R) HD Graphics 630 (1.0 GB)
GPU 2        NVIDIA Quadro M2200 (4.0 GB)

I know CPU is getting old but still it took unusually longer and became unresponsive, I thought it's hung. Anyway, will try in future after a couple of iteration.

Sample auditioning synced to tempo and key is a nice suggestion — noted.

Thank you
If you like, I can send some other feature requests. Let me know where to send (email? FR/Issue tracker?).

[pekbro]

I'm not even using this with any AI, in case some think it required, it's not...
No key, no interest in getting one tbh...

[ConceptualMachines]

Code: Select all

OS           Microsoft Windows 11 Pro (Build 26200)
CPU          Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz
RAM          32GB (16GB + 16GB) @ 2400MHz
GPU          Intel(R) HD Graphics 630 (1.0 GB)
GPU 2        NVIDIA Quadro M2200 (4.0 GB)

I know CPU is getting old but still it took unusually longer and became unresponsive, I thought it's hung. Anyway, will try in future after a couple of iteration.

Sample auditioning synced to tempo and key is a nice suggestion — noted.

Thank you
If you like, I can send some other feature requests. Let me know where to send (email? FR/Issue tracker?).

Thanks for the details. The Quadro M2200 with 4GB should actually handle a 7B fine, not blazing, but it shouldn't hang. So that sounds more like a bug than a hardware ceiling. Do you remember what you were asking it? Even a rough description helps narrow it down.
It is version 0, so bugs are expected, but I try to turn fixes around quickly 0.5.1 is already out for an unrelated localisation issue.
For feature requests and bug reports, GitHub Issues is the best place for everything: https://github.com/Conceptual-Machines/ ... ore/issues. Keeps it all in one queue.

[ConceptualMachines]

I'm not even using this with any AI, in case some think it required, it's not...
No key, no interest in getting one tbh...

Confirmed, MAGDA is first and foremost a DAW. The AI is completely optional, you can ignore it entirely and use the rest of the app as a normal DAW.
For what it's worth, I asked for the thread to be moved to this subforum because I thought the AI angle made it adjacent, but in hindsight that wasn't a great call, the AI really is just one feature among many.

[satYatunes]

Do you remember what you were asking it? Even a rough description helps narrow it down.

I do...

Code: Select all

Create a midi clip for melody in c major scale

I know it's very generic but I just wanted to try it and see what it spits out.

For feature requests and bug reports, GitHub Issues is the best place for everything: https://github.com/Conceptual-Machines/ ... ore/issues. Keeps it all in one queue.

Thank you, will use that going forward.

[ConceptualMachines]

Do you remember what you were asking it? Even a rough description helps narrow it down.

I do...

Code: Select all

Create a midi clip for melody in c major scale

I know it's very generic but I just wanted to try it and see what it spits out.

For feature requests and bug reports, GitHub Issues is the best place for everything: https://github.com/Conceptual-Machines/ ... ore/issues. Keeps it all in one queue.

Thank you, will use that going forward.

Honestly, that prompt is actually very complex. "Create a MIDI clip for melody in C major scale" may sound simple on the surface, but it isn't, even for a human. Ask a musician to actually sit down and write a convincing melody in C major from scratch and most will freeze up: which notes, in what order, for how long, over how many bars, with what rhythm, starting and ending where? The "simplicity" is an illusion that only holds until you try to produce the output yourself. And under the hood you're asking the model to handle all of that at once:

• Structural decisions: how long is the clip? What time signature? Tempo? How many bars? None of that is specified, so the model has to infer or invent it.
• Musical coherence: a melody isn't just "notes from the C major scale." It needs phrasing, contour, rhythmic variation, tension and resolution, a sense of beginning and end. Just sampling pitches from {C, D, E, F, G, A, B} gives you a random walk, not a melody.
• Rhythm generation: durations, rests, syncopation, note density — all of which need to be internally consistent and groove-aware.
• MIDI serialization: outputting valid note-on/note-off events with correct timing, velocities, and channel data in whatever format the tool expects.
• Implicit style: "melody" carries a ton of unstated assumptions (tonal center on C, cadential motion, singable range, etc.) that a human musician fills in automatically but an LLM has to reconstruct from the prompt alone.

That's why a local model struggles. It's not really a "command" — it's a generative task with musical, temporal, and structural constraints all at once. Local LLMs are fine as a router in a hybrid setup (parsing intent, dispatching to specialized tools) or for deterministic commands like "create a 4-bar empty clip", potentially you could get some results with "generate an arpeggio on Cmaj7." For actual melody generation, unless you're running serious hardware, forget about it with a local.

Maybe I should write a how-to guide for MAGDA AI covering exactly this — what to offload to a local model vs. what needs a larger one, and how to phrase prompts so you get useful output instead of noise.

[tumface]

Edit: deleted post. I see now this thread has been moved to the AI section, and I was looking at threads I had participated in the past, and didn't notice it had been moved. No need for me to reply any further.

[ConceptualMachines]

MAGDA 0.6.0 Released

Version 0.6.0 is out. This one is about hardware control:

• MIDI Controllers — Full controller stack. Right-click any parameter — macro, modulator rate, plugin parameter — pick MIDI Learn, wiggle the knob, done. Bindings follow the device across projects, not just the session. Automap pairs a controller and a device in one click using the controller profile and the device's macro layout.
• Touch & Latch automation — Completes the automation system started in 0.5.0. Touch records while you hold a control and snaps back when you release; Latch records from grab until the transport stops. Write mode is still there too.
• Modulator → Modulator links — Macros and modulators can now drive other modulators. Link a macro to an LFO's rate, or one LFO to another LFO's rate. Mod rate has its own automation lane.
• Controller-disabled indicator on every bound parameter, and uniform macro exposure across native, VST, and AU devices.
• Stability and timing fixes around clip duplicate, BPM snap, audio clip beat sync, and theme polish on JUCE-drawn title bars.

Download: magda.land
Source: GitHub
Video:

[dayvyg]

Love the interface. This looks really promising.

Are there different tools for midi input - paint, erase, cut etc.?

Thanks.

[ConceptualMachines]

MAGDA 0.7.0 Released

Version 0.7.0 is out. The three main features are Lua scripting for MIDI controllers - part 2 of the controller work that started in 0.6, presets across every device, and AI sound design for 4OSC, with audio clips, device UI, and Session View groundwork also getting significant updates:

• Lua scripting for MIDI controllers — This is part 2 of MAGDA's controller system: 0.6 added profiles and Learn, and 0.7 adds a sandboxed Lua 5.4 layer underneath the profile / Learn system from 0.6. Drop a .lua file in the scripts folder, define on_midi(event), and use the magda.* API to drive transport, session, tracks, the focused device, and MIDI out for feedback. Two-way communication, so pad grids can light up to show what's playing and DAW-protocol controllers can talk back to the host. Examples for Launchpad, Launchkey Mini MK4, and a foot-pedal template ship with the source.
• Presets across every device — Real preset menus on every device header. MAGDA presets (.mps) capture the whole device — parameters, plugin state, macros, modulators, sidechain wiring, gain. For VSTs and AUs, MAGDA scans the standard preset directories so your existing .vstpreset and .aupreset libraries appear directly in the menu. And at the track level, the entire device chain — every device, every nested rack, every modulator binding — saves as a single preset.
• AI sound design — starting with 4OSC — Devices that support it now have an AI panel docked next to the slot. 4OSC is the first one. Type a description of the sound, the model streams back a preset and applies it directly to the synth. Status line and a "starting point only" reminder, plus the AI's chosen preset name and category auto-fill the save dialog so committing a generated patch into the new preset system is two clicks.
• Audio clip redesign — New waveform painting with a peak cache, smoother scrolling on long files, clearer clip-selection strip and time-range colour inversion. Each audio clip now carries its own source BPM, so beat mode and warp stay consistent across project tempo changes.
• Device redesign — Volume slider moved out of the device header and onto the right, sitting right above the meter where it belongs. Right-edge icon order (preset / power / delete) is now structurally locked.
• Session view groundwork — Multi-clip selection, per-slot play and stop, and an optional launch fade between clips. Most of this is foundation for session recording, which lands in the next release.
• Plus configurable user-data folders, multi-track selection, AI-chat @ aliases for focused-device parameters and macros, and a stack of bug fixes (session mini-mixer clipping, beat-mode BPM caching, clip loop-range undo, more).

Download: magda.land
Source: GitHub
Video:

[ConceptualMachines]

Love the interface. This looks really promising.

Are there different tools for midi input - paint, erase, cut etc.?

Thanks.

Thanks, appreciate it!

No tools yet, I tend to prefer key modifiers but I can see the appeal. I’ll probably revisit it once the core features are in.