How to Use Lyria 3 Pro with a Signed-In Browser

Lyria 3 Pro cannot read a logged-in web page on its own. It is a music model, not a browser agent: Google’s model documentation lists no function calling, URL context, or search grounding. To turn an approved music page into a usable prompt, use a local text agent with Tabbit CLI to read the page in Tabbit Browser, then review the draft before opening a separate Google music-generation surface. The model ID is lyria-3-pro-preview; Google currently marks it Preview and recommends lyria-3.5 as its successor, with no Pro shutdown date announced.

Install Tabbit Browser and Tabbit CLI on the same machine, then install the Tabbit skill in your local agent client using that client’s documented setup.

This separation matters. A model appearing in a menu does not prove your account can run it, and Tabbit’s local browser connection does not provide a music-generation entitlement. Check both independently.

A Reddit user who tried Pro said it reproduced a requested “guitar fades in at the intro” cue, while finding aggressive genres too safe. A different user comparing earlier Lyria 3 / Producer.ai said its voices sounded more natural to them. That second post is not a Pro or Clip test, and neither report evaluates this browser workflow. Pro report · Lyria 3 comparison.

Key takeaways

• Lyria 3 Pro generates audio; a separate local text agent handles the signed-in browser page.

• Use lyria-3-pro-preview only after confirming that your selected Google project and model selector expose it.

• Google Cloud lists text or image input, MP3 output, one clip per prompt, and a 184-second maximum.

• Install Tabbit Browser and the Tabbit CLI skill in a local client. Pi uses /skill:tabbit; other clients may show /skills or a client-specific skill picker.

• Keep the flow read-only through page review. Approve the prompt yourself and generate audio in a separate Google surface.

Which component does what?

Google Cloud documents lyria-3-pro-preview as a Preview model with text and image input, audio output, and a 184-second maximum. It lists BPM, intensity, and duration controls, but does not support negative prompting. Its page says one clip per prompt and names MP3 at 44.1 kHz. The Gemini API lifecycle page lists lyria-3.5 as the recommended replacement for Pro and no shutdown date for lyria-3-pro-preview. Google’s Pro model page currently displays the Clip ID in its model-code row, so use the Cloud model table and lifecycle page to verify the ID rather than copying that inconsistent field. Google Cloud model details · Gemini API deprecations.

The model is not an agent host. The Google model page explicitly reports no function calling or URL context. If you want a text model to inspect the browser page, check the client’s model picker, provider credentials, and local tool permissions first. The fact that an agent can discover a tabbit skill does not prove its launcher can reach the browser; a working CLI does not prove a website login; and neither check proves Lyria access. The Tabbit CLI guide and browser MCP versus CLI guide explain the local connection choices.

Prepare the approved page without changing it

1. Open Tabbit Browser and sign in yourself to the music brief, project page, or review board. Confirm the page title and approval status in the intended profile.

2. In a local client, check its skill interface. Use /tabbit only if that client documents that shortcut. Pi’s documented command is /skill:tabbit; some clients instead expose installed skills through /skills.

3. Run the installed CLI’s diagnose command. Confirm the local launcher connects before asking the model to browse.

4. Ask the agent to read only the approved title, status, audience, mood, and allowed assets. Do not copy confidential source material or personal data into an external music prompt.

5. Keep your ordinary research tab open. The CLI task uses a separate task tab; you can compare its short evidence report with the visible page. See how to connect an agent to an existing browser and the first-task checklist.

Tabbit’s existing login is tied to the selected Tabbit Browser environment. The CLI does not sign you in or ask for your password. If the task lands on a login page, stop and sign in manually. A local fixture test recorded in this worktree showed one protected page returning 401 before a test session was set and a separate CLI task reading the same fixture after sign-in. That check supports session reuse in that fixture only; it did not run Lyria or prove a model-to-browser integration.

Draft a Pro prompt from the page

Ask the local text agent to return a brief, not to submit a generation request. This example is a template, not a tested Lyria output. It requests a medium-length instrumental cue and leaves final choices to the human reviewer.

Use Tabbit CLI in the current Tabbit Browser profile. Read the page I have open.
If its status is not Approved, stop and report the visible title and status.
Do not click approval, publish, upload, or change any fields.

For an Approved brief, return: page title, visible status, intended listener,
scene, emotional direction, permitted references, and any missing decision.
Then draft one Lyria 3 Pro prompt for a 90-second instrumental cue.
Specify genre in broad terms, tempo range, key instruments, energy, and a simple
intro / build / ending. Do not name artists, quote lyrics, or invent facts.
Keep the source URL beside the draft and stop for my review.

Expected output: a short evidence block followed by one prompt in a code block.
If the page does not show approval, return no generation prompt.

The prompt names a target duration but does not guarantee an exact result. Lyria’s technical maximum is 184 seconds; its outputs can vary. Google documents a watermark and content filtering for Lyria. Review the generated preview for fit, lyric or likeness concerns, and the applicable terms before you use it. See Google’s Lyria model card and the service-specific Google Cloud Preview terms.

Review before any sharing

If Pro is absent from your current selector, do not route a prompt through an unknown alias. Google’s deprecation page points to lyria-3.5; confirm its availability and terms in the service you plan to use, or leave the draft at the prompt stage. For a broader overview of the local session boundary, read reusing a signed-in browser session and researching with browser sources.

The useful boundary

Use Tabbit Browser when the source of the creative brief sits behind your login and you want a local agent to extract only a few approved fields. Use Lyria in its own supported Google surface to generate the audio. Tabbit’s task tab can stay separate while you continue reviewing the brief in your regular tab. A browser agent should stop at the prompt and evidence block; you decide whether to generate, download, or publish. For CMS review practices, see drafting in a CMS with a browser agent and validating a local browser task.

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.