Studio MCP Server
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Studio MCP Serveradd a wobbly sub bass at 160 BPM and play the song"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
AgentLoops
FL Studio, but for AI agents. A song is a multi-agent pipeline you can see, hear and edit like a beat. Channels are agents, notes are agent calls, patterns loop, mixer buses carry each agent's output to the next one, and what reaches Master is the result. A song is also music: every channel has an instrument, so a song plays as you build it and exports to WAV or MP3.
Use it from Claude Code, Codex, Cursor, opencode or any MCP client. The coding agent drives the studio through the Studio MCP server, and you watch each edit land in the studio window. You don't need an API key. script channels run locally, and harness channels hand their notes to your coding agent. Only llm.agent channels call the Claude API.

Requirements
Node.js 24 or newer.
A Chromium-family browser (Chromium, Chrome, Brave or Edge) for the app window and for exporting audio with no studio open. Without one, the studio opens in your default browser.
ffmpeg, optional, for MP3 export. WAV needs nothing extra.
Related MCP server: opendaw-mcp
Install
npm install -g @nawka12/agentloopsThat gives you the agentloops command. To try it without installing, use npx @nawka12/agentloops <command>.
Straight from GitHub:
npm install -g --allow-git=root github:nawka12/agentloopsnpm 12 refuses git installs unless you allow them, and --allow-git=root allows only this one. If npm then warns that the package's prepare script was blocked, you can ignore it: the package was already built while it was installed.
From a clone:
npm install # also builds dist/
npm linkYou can also skip the install and run node src/cli.ts <command> from the clone.
Make a song with Claude Code
agentloops init my-song --harness claude-code # a starter song, .mcp.json and the AgentLoops skill
cd my-song
agentloops studio song.alp.json # opens the studio in its own windowIn a second terminal, in the same folder:
claude # approve the "agentloops" MCP server when askedThen ask for what you want:
Make a kawaii future-bass song at 160 BPM: a supersaw chord stab, a wobbly sub bass, chip arps and a lyric line. 16 bars with an intro, a build and a drop. Play it for me, then export it as an MP3.
The agent reads the manual from its skill, writes the song, plays it and exports the MP3. Every change shows up in the studio, labeled with who made it, and Ctrl+Z undoes any of them. You can keep editing in the studio at the same time: the agent sees your changes on its next read.
Other harnesses use the same setup with a different --harness:
| Writes | Then |
|
| run |
|
| run |
|
| turn the server on in Settings → Tools & MCP |
|
| run |
|
|
The studio's ⇄ connect button writes the same files for the song that is open.
Try the demo
agentloops studio "$(npm root -g)/@nawka12/agentloops/examples/kawaii-bass.alp.json"Press Space to play. Shift+click ⤓ saves an MP3 next to the song. It is 52 bars, with 8 script channels for chords, bass, arp, bell hook, drums, riser, vocal chops and a lyricist. The lyricist listens to the vocal chops and sends each sung line to Master, so the Monitor (F11) shows the lyrics as they're sung.
Commands
Command | What it does |
| Create a song folder wired to a harness |
| Open the studio. With no file it opens |
| The Studio MCP server on stdio, which harnesses start for you. It relays to an open studio or runs headless |
| Check a song and explain each error with a fix |
| Run the song with no UI and print what reached Master |
| Print the arrangement and patterns as a text step grid |
| Linux: add AgentLoops to the app menu and open |
How it works
The studio is the host.
agentloops studioholds the song, the undo history and the runs.agentloops mcprelays every tool call to the open studio over localhost, using a token the studio writes to.agentloops/studio.json. With no studio open, it hosts the song itself.Three scopes.
readcan only look.editcan change the song, and every change is saved and undoable.runcan also play the song, which spends money if it hasllm.agentchannels, so runs are capped by the song's budgets and the--budgetthe harness was started with.Time is barriers. Notes at the same position run in parallel, and a note starts only after every note ending at or before it has finished. Playback tempo is a speed limit, so a slow agent shows up as a held note.
Takes are cached. Every run is saved as a take in
.agentloops/takes.sqlite. A voice whose prompt and inputs did not change replays from the cache for free, which is what makes looping a pattern cheap.Generators.
scriptruns JavaScript in a sandbox.llm.agentcalls Claude with tools and MCP servers.harnesshands a note to the connected coding agent: the run pauses until the agent answers withvoice_reply.
Docs
docs/manual.md: the manual for agents and people, covering the project format, every MCP tool and how to make music.
initcopies it into the skill orAGENTS.md.SPEC.md: the full design, with the FL Studio research behind it.
docs/case-study-blender-mcp.md: why the studio is the MCP host.
This server cannot be deployed
Maintenance
Related MCP Connectors
Autonomous music production for AI agents with MIDI generation, QC and provenance.
Generate full songs with vocals, instrumentals and sound effects from text. One server for Suno, Google Lyria, Mureka, MiniMax, ACE-Step, YuE and ElevenLabs. Your agent can quote the credit cost before generating, start a song, wait for it to finish and return download links. Sign in with OAuth; optional daily credit caps; retries never charge twice.
AI music and podcast platform for autonomous agents. SoundCloud for AI bots.
Music network for AI agents: get a song from your name, post, review tracks, leave notes.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to control Ableton Live with full access to Live's object model, including clip creation, device control, automation, and audio signal capture for mixing and mastering tasks.1MIT
- AlicenseBqualityAmaintenanceEnables AI agents to control a browser-based digital audio workstation (openDAW) for music production, including track creation, effects, MIDI, automation, and rendering.100Apache 2.0
- AlicenseBqualityDmaintenanceEnables coding agents to compose, tune, render, mix, and audit native VOCALOID3/4 projects from scratch, acting as a production bridge between intent and finished song.2214MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to create and edit Strudel music code, render offline WAV audio, and obtain structured hearing reports with waveform, spectrogram, BPM, and onset analysis.AGPL 3.0