draft 2026-10-20
last checked 2026-10-22
ismail vs SoundCraft: which one should your agent use?
Last checked: 22 October 2026. SoundCraft is pre-alpha and changes daily; every line about it links to where we read it.
Short answer
If you record and edit audio the Pro Tools way, want your plugins and MIDI, and want an agent to drive the same app you use, use SoundCraft. If you want the agent to make the music itself and check every render as text before you listen, use ismail. They do different jobs and can work together through audio files.
What each one is
- SoundCraft is an open-source, clean-room re-implementation of Avid Pro Tools in Rust, for macOS, Windows, Linux, FreeBSD and the browser (README). Every menu action is a command with JSON parameters, and an MCP server lets an agent list commands, run them (up to 1,000 in a batch), inspect the session, bounce the mix and take screenshots of the app (MCP docs). Its README calls it pre-alpha.
- ismail is a DAW written for an agent. The agent writes notes, instruments, effects and the mix as text, ismail renders the audio, and the agent gets every render back as text: levels per bar, drum lanes, piano rolls, chords, song structure and scored comparisons against a reference. The agent also builds the screens you steer it through: a phone page, feedback pages, a VR stage.
Side by side
| ismail | SoundCraft | |
|---|---|---|
| Built for | An agent making the music, with a person listening and steering | A person editing and mixing, with an agent able to drive every menu |
| A screen for you | Built by the agent when you need it: a phone page, feedback pages, a VR stage | A full Pro Tools-style desktop app |
| What the agent gets back about the sound | Every render as text: about 40 tools for levels, grid, key, chords, melody, drums, swing, structure, timbre and comparisons | Peak and integrated loudness on a bounce, the session state and screenshots (MCP docs); beat times and tempo found in audio (Beat Detective), and a chord name per bar from MIDI notes (Extract Chords) (code) |
| Instruments | Its own synths, drum synths, samplers, voices written in Python, and instruments measured from a recording | Its own synths, drum synth and sampler, plus CLAP, VST3 and AU plugins (README) |
| Recording | A person's take goes in as an audio file; a live set's output can be recorded | Multitrack recording and comping (README) |
| MIDI | Live MIDI in through a live set; no MIDI files in or out | MIDI files and MIDI out (README) |
| Matching a reference | Grid and swing measured from a record, scored comparisons, transcription by consensus | Beat Detective finds beats in audio; Extract Chords names chords from MIDI notes (code) |
| Playing live | Decks, transitions and edits while the music runs | Not described in its docs |
| Tempo | One tempo per project (a tempo change is in review) | Tempo maps (README) |
| Surround | Stereo | Up to 9.1.6 and Ambisonics (README) |
| The song is | A folder of text files the agent writes directly: notes, voices, effects and the mix | A JSON session file (.scraft) with an Audio Files folder (file formats) |
| Licence and cost | Free, MIT | Free, open source (Apache-2.0 per GitHub) |
| GitHub stars (22 Oct 2026) | 17 | 3,533 |
What "the render comes back as text" looks like
This is part of what ismail's agent reads from a finished track, Uke Story. One character per bar (9 = loudest, each step down = 4 dB):
bar 123456789012345678901234567890123456789012345678901234567890...
mix 66666687787787787887888888888888888888888888888878889999888888...
~sub 777777777777777777777776777777666666777777777777.7899999999999...
section A----------------------------------------------|B------|C------...
A bars 1-48 (48) level 9.6 dB low X...X...X...X... snare ....x.......x...
B bars 49-56 ( 8) level 12.6 dB low XxxXX.XxXxxXxxXx
C bars 57-78 (22) level 12.7 dB low XxXXXxXxXXXX.XXx snare x.xXx..xX.x.x..x
The agent can see where the sub drops out, which bars repeat and where the kick pattern changes. In SoundCraft, an agent can get the loudness and peak of a bounce, beat times from audio and chord names from MIDI notes. ismail's readings cover the whole mix: levels per band and bar, sections, drum patterns, and comparisons against a reference.
Choose SoundCraft if
- You record, comp and edit audio and want a Pro Tools workflow without Pro Tools.
- You rely on your plugins, MIDI files or surround.
- You want to stay at the controls and let an agent handle the tedious edits in the same app.
Choose ismail if
- You want the agent to do the whole job: write, render, check the result, fix, render again, and ask you what you hear.
- You want to steer by listening and talking, from your phone or a VR stage, not by clicking through menus.
- You want to build an instrument from a few recorded notes, or match the feel of a reference record.
- You want the agent to play a set live.
Can I use both?
Yes, through audio files. ismail renders each track as its own WAV stem, which a DAW can import. ismail can measure and plan; SoundCraft can host your plugins and edit recordings.
Where ismail falls short
ismail says this on its own README: a new song from a blank page can sound generic on the first sketch, and it is strongest with something to measure. It has no singing voice, one tempo per project, no MIDI files and no plugin hosting. It needs a capable agent, and your ear stays the final judge.
Try it
Paste this to your agent (Claude Code, Codex, Cursor, or the Code tab in the Claude desktop app):
Set me up with ismail: https://github.com/newsbubbles/ismail
Sources, read 20 October 2026 and re-checked 22 October 2026: SoundCraft README, docs/mcp.md, docs/control-protocol.md and its menu catalog; star counts and licence from the GitHub API. ismail facts from its README and its tool list on main.