Make Instrumental Music with an AI Agent

An AI agent can turn a written music brief into a track, then revise it when you describe what to change. An agent is software that works toward a goal by understanding instructions and using tools. In music-making, those tools can arrange notes or create sounds electronically. Instrumental music has no sung vocals, though a track can still contain voice-like sounds if the tools allow them.

An agent writes a score, a file of musical instructions, and Soundroo turns the score into audio. You describe the style and purpose, listen to the result, and ask for revisions such as “faster,” “less piano,” or “keep the bass line.” The agent handles the musical instructions, while you decide whether the track works.

How an agent makes an instrumental track

A music agent can interpret your request, create musical parts, and use a music tool to produce sound. Research on MusicAgent describes this kind of setup. An AI program trained to work with language plans tasks and calls specialized tools.

Soundroo represents each track with a score, a structured text file in JSON format that lists its tempo and instruments, along with notes, patterns, and sections. Soundroo’s JavaScript engine creates sounds electronically from scratch and mixes them into stereo audio. It uses no music AI model or instrument recordings, and it renders the same score the same way each time.

The score can show a drum pattern as a row of symbols: X marks a loud hit, x a normal hit, o a soft hit, and . a rest. A melody can use note names and durations, such as G4/.5 Bb4 C5/1. You do not need to write these by hand: the agent creates and edits the score.

Soundroo includes piano, electric guitar, bass, and drums. It also offers synth leads, pads, strings, bells, and marimba. Piano and strings are synthesized approximations. Soundroo cannot sing, and it does not make sung vocals. Short spoken phrases can become rhythmic chops, textures, or a vocoder-style robot voice.

Give the agent a useful brief

A clear brief names the track’s purpose, mood, length, and instruments. It also says how the music should behave around other audio. For example, background music under narration should leave space for the speaker.

Paste this prompt into an AI agent and replace the bracketed text:

Make me a track: [what it should sound like, or "surprise me"]. Use Soundroo for it. If you have Soundroo tools, use them. If you can run commands, clone https://github.com/vincentsch/soundroo, read its AGENTS.md and follow the steps. If you can't do either, send me to https://soundroo.com/connect

For example, describe “a warm, curious instrumental with soft piano, muted guitar, and light drums, leaving room for spoken narration.” Then listen and give targeted feedback. “Make the opening less busy” is easier to act on than changing the mood, tempo, instruments, and structure all at once.

The agent cannot hear music as a person does. It checks the audio for technical problems such as clipping, silence, too much limiting (audio processing that reduces loud peaks), and sudden changes in volume. You still need to listen for musical fit and tell the agent what to revise. Older versions remain available as it makes changes.

Use Soundroo with an agent

Soundroo is free and open source under the MIT license. The way you connect it depends on the agent:

• Coding agents that can run commands, such as Claude Code and Codex, use the Soundroo toolkit on your computer. This route needs Node.js, but making music does not require an account, API key, music software, GPU, or internet connection.
• ChatGPT and Claude connect through Soundroo’s MCP server, an online service that lets them use Soundroo tools. MCP is a standard that lets an assistant use an external tool. Add the Soundroo connector once, then sign in with Google, GitHub, or an email link.

Chat apps make shorter, simpler pieces, typically one to three per conversation, and return a private preview link. The link expires after a day unless you publish the track. For a guided starting point, the custom prompt builder lets you set styles, track count, length, and spoken words.

The coding toolkit also opens a listening studio in your browser, with a player and generated cover art. It also includes favorites, notes tied to moments in the track, a mixer, and older versions. You can export WAV audio, the score, MIDI, or instrument stems. You can also export an offline listening page. MIDI contains instructions for notes and musical events; stems are separate audio parts, such as drums or bass.

Listen, revise, and decide whether to share

Start by checking whether the track fits its purpose and whether its opening and ending work. Also listen for voice-like sounds that distract from the instrumental. Then ask for one or two specific changes and listen to the new version. Since the agent cannot judge the music by listening, your feedback guides the creative revisions.

Sharing on soundroo.com is optional. On the first share, Soundroo creates an unlisted artist page with a generated name; anyone with the link can listen. You can keep the page by confirming an email address. Shared tracks can be open, listen only, public, unlisted, or private. Open tracks are downloadable and remixable with credit under a Creative Commons Attribution license. Soundroo refuses exact copies of other artists’ tracks.

Soundroo is free, while the AI agent you use is a separate service with its own pricing. Check the agent’s terms before using it, and choose Soundroo’s sharing and download settings to match your plans.