Music Generation MCP Server

「セカンドチャンス」

Second Chance

Initially, I used Suno to create the theme song for Reincarnator x Regressor. While I composed the lyrics, the restrictions grated on me. As a free user, I had limited downloads, which meant limited iterations. Apparently, there's a time constraint too. I also didn't own the full song because I didn't have a paid subscription. Make no mistake, Suno is great, but for someone with a limited budget, and considering that my promise to future users was that they wouldn't require additional subscriptions, and could do all this on a single AI subscription, I decided to build a local music generation MCP server.

That would sidestep any licensing issues (even if the music isn't as good as Suno — while I was fairly satisfied with Suno's first attempt, I spent six rounds of listening across nineteen takes before finally settling on this) and additional monetary costs. But you get what you pay for...only, this time, at least you own what you produce without having to pay extra for it.

ACE-Step, through ComfyUI

ACE-Step is the only credible open model that has vocals. MusicGen and Stable Audio Open are instrumental-only, which is fine for background music. Unfortunately, given that I composed my own lyrics, I didn't want my efforts to go to waste, so ACE-Step was my only real option. It runs through ComfyUI, which has supported it natively since v0.25, so there's nothing to install but the model weights. That made it a materially easier setup than the local video path, which needed third-party loaders.

The server takes style tags, lyrics, tempo, key, time signature and language, and returns a lossless FLAC to archive plus a 320k MP3 the Remotion pipeline can drop straight in. It also emits a beat grid, so trailer cuts land on downbeats rather than wherever the edit happened to fall. Claude Code seems to think this is a great idea for making videos come to life, so I decided to let him test that out.

This also allowed me to not worry about iterating as many times as I wanted. Not only could I iterate dozens of tracks within the same 5-hour session (admittedly, I'm currently subscribed to the Max plan of Claude Code, so I don't know if it'll burn through a Pro user's 5-hour limit), they were produced in an impressively short amount of time. I think Claude, using my MCP server, produced a 107-second song in about a minute and a half. And that's when he's working on my 6GB VRAM laptop.

The interesting part wasn't the generation — it was working out what the controls do

These are Claude Code's observations from running the sweeps, not mine — I was the one listening and deciding, he was the one keeping track of what the settings actually did.

• Lyric density is arithmetic, and it's the master control. One bar is 4 × 60 / bpm seconds. Divide the track's bars by its sung lines: land near 2.3 bars per line and lines get sung as lines. At 4.7 they stretch into fragments with gaps. Below 2 they run together, and trailing lines get silently dropped.
• Instrumental sections aren't free duration. They compete with the lyric for the same bars. Adding an intro, three interludes and an outro to an 85-second track quietly truncated the song at line 10 of 16.
• Asking for a relative minor does nothing. I picked B minor to keep the mood "related" to the D major I had. B minor is D major's seven notes — that's what a relative pair is. Sharing a pitch collection is exactly what made it unable to sound less bright.
• Some things aren't controllable at all. Vocal register ignores the prompt: five takes at identical settings returned one female vocal and one noticeably high-pitched, despite explicit deep male baritone tags. Sample and pick; don't tune.

Infinite iterations...at least, as infinite as your current AI subscription allows

Music is generated locally, at no additional cost (beyond your current AI subscription), and unlimited, with the caveat that it's still constrained by your AI subscription. Obviously — you didn't think generating music wouldn't need tokens, did you?

However, this is still a game-changer. Producing almost twenty iterations of a song across a couple of hours without worrying about my 5-hour limit (again, this is on Max plan, so I can't promise anything for Pro users) or needing to subscribe to a dedicated music AI generator? I'll certainly take that.

What's new

Earlier releases

A growing Webcomic MCP network

This is the newest piece of a wider webcomic and animation toolkit: the Background Generator for backgrounds, the Character & Panel Generator for consistent characters across a webcomic, Novel Translation for translating my novel into other languages, and Anime Production for producing animated videos. In my roadmap, I still have a speech bubble server still to come.

With the Music Generator MCP server, I would like to think I have developed a self-contained, living ecosystem that allows you to do everything on a single AI client. Instead of having to subscribe to multiple AI generators, such as Midjourney, Kling and Suno, and still needing to compile all of the disparate forms of media into a single video or product, I can do it all on Claude Code (and you on your single AI client, whether it's Codex, Antigravity, Kimi or Cursor). That's the objective I set out to achieve, and as more possibilities open up in the future, I'll try to add whatever webcomic artists and web novelists might need to make online storytelling more accessible and efficient.