Create Songs on Kubeez: UI Walkthrough + MCP Automation

Music used to be the slowest step in a content pipeline — licensing, stock libraries, endless re-bouncing. Kubeez replaces that: original, Suno-class tracks generated from a short brief, with full stems, add-vocals, and lyrics tooling built in. This post walks through the UI first — the path most creators take — and then shows how to automate the same workflow through the Kubeez MCP server so any MCP-compatible assistant (Claude, Cursor, ChatGPT, custom agents) can hand you finished tracks without a browser tab.

#What Kubeez music can actually do

Before the walkthrough, the short version of the surface area:

• Full songs with vocals or instrumentals only, at multiple durations.
• Lyrics generation — ask for a theme, get a verse/chorus structure, then generate over those lyrics.
• Add vocals to an existing instrumental track.
• Stem separation — split any generated or uploaded track into vocals / drums / bass / other.
• Commercial-ready exports — watermark-free, no attribution required.

That matters because "AI music" is often sold as the demo and none of the polish. Kubeez covers the whole pipeline from brief to final-stem multitrack.

#UI walkthrough: your first song in under two minutes

1. Open the Music generation page. The default view is a clean prompt box plus a few starter styles — think dramatic symphony, pop song, rap instrumental. Click any of them to see a working prompt you can edit.
2. Write the brief. Specific wins: "Warm indie-folk track, acoustic guitar and light percussion, mellow male vocals, about late autumn light and a drive home, 90 BPM, 2 minutes." Vague loses: "a sad song." Kubeez passes your text straight to the model — the richer the brief, the closer the first take lands.
3. Pick vocals or instrumental. The toggle sits next to the prompt. Instrumentals cost the same but let you drop your own vocal on top later with Add Vocals.
4. Optionally draft lyrics first. If you have a theme but not lines, generate lyrics first, then paste them back into the music generation prompt. Kubeez keeps the vocal take tightly aligned to your lyrics instead of hallucinating syllables.
5. Generate. Kubeez queues the job and shows progress. Music generations typically take 40–90 seconds end-to-end.
6. Preview in the waveform player. You can scrub, loop sections, and A/B between variants without leaving the page.
7. Export or keep iterating. Every generation saves to your history; if a take isn't quite right, tweak the brief and regenerate — your credit balance and the prior versions stay visible the whole time.

#After the first take — what pros do next

One generation is a starting point, not a finish line. The creators getting real mileage out of Kubeez tend to:

• Generate with vocals → split to stems so they can swap the vocal take for a custom one later.
• Generate an instrumental → Add Vocals with a different lyric set for the same campaign (same bed, two language variants, three artist-voice options).
• Export + auto-caption the song as part of a short-form video — Kubeez does both in the same account.
• Keep a running brief doc — pasteable paragraphs for "our sound" that they drop into each new generation for consistency.

The UI is optimized for the first two takes of a creative flow. Once you know the shape of what you want, automation starts paying off.

#Automating with MCP — any compatible agent

The Model Context Protocol (MCP) is an open standard for connecting AI assistants to tools, prompts, and resources. Kubeez's MCP server exposes the same music surface you just used in the UI — generate_music, generate_dialogue, get_music_status, generate_separation, get_separation_status, plus balance and model lookups.

Any MCP-compatible client can drive it:

• Claude Desktop and Claude Code — OAuth sign-in, first-class tool call UI.
• Cursor — add Kubeez via the MCP server URL, same tools, in-IDE chat.
• ChatGPT (MCP-supporting builds) and other hosts that accept a remote MCP endpoint.
• Your own agent — any runtime with an MCP client library.

The authoritative reference is the Model Context Protocol specification. Setup guides for Kubeez specifically live in MCP overview and Quick start.

#Setup in under a minute

1. Open MCP settings in your Kubeez account. Copy the server URL: https://mcp.kubeez.com/mcp.
2. Add Kubeez as an MCP server in your client. OAuth is recommended — the client opens a browser, you approve, no secret is pasted. If your client only accepts a static token, generate a personal access token from the same page and use it as Authorization: Bearer ….
3. Reload the MCP connection until the music tools appear alongside the media ones.

#What a music-focused agent brief looks like

A concrete, reusable prompt you can paste into any MCP-connected assistant:

"I'm running a 30-day audio sprint for a meditation app. Every week I need 3 distinct ambient tracks, 2–3 minutes each, instrumental only, labeled by mood (calm, focus, sleep). Check my balance before each batch. If an individual generation fails, retry once with a slightly different brief. When all three land, give me the CDN URLs and a one-line description of each."

What the assistant typically does:

1. get_balance — confirms the budget is available.
2. get_models filtered to model_type=music — picks the right model and quotes the per-track cost.
3. For each track: generate_music with the brief → get_music_status polling every 10 seconds until completed.
4. On failure: reads the error message, adjusts the brief, retries once.
5. Summarizes: three CDN URLs, moods tagged, credits spent.

No dashboard tabs. No prompt copy-paste. If you live in Claude or Cursor, the tracks land in your chat and your credits are the audit log.

#Stem separation, add-vocals, and longer workflows

Once you're comfortable with generate_music, the rest of the music toolchain slots in the same way:

• Add Vocals — upload (or re-use a generated) instrumental, provide lyrics, and get a vocal take mixed over it. Great for multi-language versions of the same song.
• Stem Separation — pass any track (yours or generated) to generate_separation to split into vocals / drums / bass / other. Poll get_separation_status for the individual stem URLs.
• Auto-captions over music videos — pair a generated song with a generated video and Kubeez will burn subtitles on top.

The MCP server exposes each of these as a first-class tool. Claude or Cursor can chain them: "Generate a lo-fi instrumental, then add my lyrics as a male vocal, then split the result into stems for post-production."

#Best practices

• Start with the brief in natural language. Genre, instruments, mood, tempo (BPM), vocal gender if any, a one-line story. Short abstract words like "vibes" produce bland results.
• Keep a reusable style paragraph. A two-sentence description of your sound (palette, influences, don'ts) that you paste into every new generation keeps the catalog consistent.
• Pick instrumentals when you'll use Add Vocals. You get the same bed with controllable lyrics per variant.
• Budget batch generations. Tell your assistant "stay under X credits for this run" — it'll call get_balance and get_models to fit the plan.
• Save takes you like. Kubeez keeps your history, but export the MP3 URLs for archival — models and prices change.

#Security and hygiene

• Revoke any personal access token you stop using from MCP settings.
• Prefer OAuth for interactive hosts — nothing to leak into a log file.
• Never commit a PAT to git. Store it in your host's credential manager.

#Where to go next

• In the product: Music generation, Audio hub, Auto-captions.
• Docs: MCP overview, Quick start, REST API.
• Related reads: Automate AI media: REST API vs MCP, Claude Design, but for media, AI models guide.

Summary: Kubeez turns a short brief into original music — with or without vocals, lyrics-aligned, stem-separable, commercial-ready. The UI at /music is the fastest way to make the first track; the MCP server at https://mcp.kubeez.com/mcp lets any MCP-compatible agent (Claude, Cursor, ChatGPT, custom runtimes) drive generate_music, generate_separation, get_music_status end-to-end without opening a browser.

Next steps

• Finish OAuth in MCP settings.
• Make your first song from the UI at /music.
• Then ask your MCP-connected assistant for a batch: "Generate three 2-minute indie-folk instrumentals, different moods, stay under 100 credits."