New 'ismail' DAW lets AI agents compose music via code

Overview

A new open-source project named ismail has emerged as a specialized Digital Audio Workstation (DAW) designed specifically for operation by artificial intelligence agents. Unlike traditional music production software built for human users with visual interfaces, ismail operates entirely through text-based instructions, allowing AI models to compose, render, and analyze music without requiring auditory perception.

The project, developed by GitHub user newsbubbles, represents a significant shift in how AI handles creative tasks. Rather than generating audio directly from prompts like services such as Suno, ismail provides a toolkit that enables an agent to write songs as code, render them, and iteratively refine the output through precise textual feedback.

ismail is not a model that turns a prompt into audio. It is a set of tools your own agent uses to write the song as notes, sounds and code, render it, read back what came out, and edit it.

This approach prioritizes precision and reproducibility, allowing developers to treat musical compositions with the same version control and structural rigor applied to software engineering.

How It Works

At its core, ismail functions as a bridge between natural language processing and audio synthesis. Every aspect of a music project—notes, instrument patches, effect chains, and automation data—is input as text. Conversely, the system returns all analysis and feedback in text format, including levels, spectra, drum patterns, piano rolls, chords, and song structures.

The system includes tools for comparing a generated track against a reference track, enabling the AI agent to judge quality and similarity without "hearing" the music in the traditional sense. While spectrogram images are available, they are optional, ensuring the workflow remains lightweight and focused on textual data.

One of the standout features is the ability to perform iterative, precise edits. If a user wishes to change a single note, adjust one synthesizer patch, or alter a specific bar, the agent can do so and re-render the track. This process leaves all other elements untouched, a level of surgical precision that is often difficult to achieve in conventional DAWs where visual selection can sometimes lead to unintended changes.

The project files are essentially code. Each song comes with a project file and a build script, meaning standard developer workflows such as git history tracking, diffs, branching, and code review are fully applicable to music production. This invites a new paradigm where musical compositions can be reviewed, forked, and improved collaboratively, much like software repositories.

Supported Tools and Integration

ismail is distributed under the MIT license, emphasizing its open-source nature. It runs locally on Windows, macOS, and Linux, ensuring users have full control over their data and processing without relying on cloud subscriptions or content filters. The system is designed to improve alongside the AI model using it; because the music is the agent's own work, a more capable model utilizing the same tools should theoretically produce higher-quality compositions.

The platform supports a wide range of sound sources, including synthesizers, drum synths, samplers, and even custom voices written in Python. This flexibility allows any sound to become an instrument within the agent's toolkit.

For integration with AI coding assistants, ismail offers multiple entry points:

1. MCP Server: It includes a Model Context Protocol server compatible with tools like Claude Code and Cursor. Running python -m ismail.mcp_server exposes approximately 70 tools over stdio.
2. CLI: Users can execute operations via the command line using python -m ismail -p <project> <op> [args].
3. Python API: Developers can import the library directly into their scripts for custom automation.

The project also includes a recommended "skill" file that teaches agents best practices, such as planning a session sheet before writing notes and writing a listening report after every render. This guidance helps ensure the AI agent approaches composition with a structured methodology rather than random generation.

Comparison with Existing Solutions

The developers of ismail acknowledge that the tool serves a different purpose than popular generative audio models. Services like Suno excel at creating polished, realistic songs from simple text prompts in under a minute, particularly for tasks requiring realistic sung vocals and genre styles learned from extensive recorded music libraries.

However, ismail argues that existing DAWs like Ableton or FL Studio are ill-suited for AI agents because they were built for humans with ears and mice. While agents can control these programs through automation bridges, they cannot effectively perceive the results of their actions. ismail solves this by putting everything an agent needs to write and perceive into compact text.

Sample tracks created with ismail demonstrate its capabilities, including a dubstep track with synthesized bass, a Chopin-inspired piece with no samples, and a drum and bass track at 174 BPM. These examples highlight the system's ability to handle complex genres and precise timing through code-driven composition.

Installation requires Python 3.10 or newer and involves cloning the repository and installing dependencies. Optional packages are available for perceptual metrics and stem separation, though users may need to handle conflicts with tools like demucs manually. For vocal tracks, the system utilizes the native text-to-speech engines of the operating system, such as SAPI on Windows, say on macOS, and espeak-ng on Linux.

As AI agents become more sophisticated, tools like ismail provide the infrastructure needed for them to engage in creative processes with the same granularity and control previously reserved for human composers and engineers.