Skip to content
voicefieldguide

How to use ElevenLabs voices from Claude

Updated November 6, 2026

The picks

Press play to hear each voice read this guide's sample line.

These picks compare my fifteen voices. I own them, and ElevenLabs may pay me when you use them.

Jett·Short demo narrator

Brisk explanation for a script written by an agent. Padding is easy to hear, so edit before generating.

0:10

Dexter·Agent-written walkthrough

Instructional cadence suits a useful demonstration. Excited product-launch copy needs a different performance.

0:11

Titan·Serious scripted narration

Weight and clear consonants suit a factual explainer. A casual chat summary may sound too solemn.

0:11
Add Jett to My Voices

Use ElevenLabs voices from Claude by giving it a casting reference and, if you want audio generation, a connected ElevenLabs tool or local script. My Prestige Voices skill helps choose among my fifteen voices and write for the chosen narrator. A connection to your own ElevenLabs account supplies generation access. I’d cast Jett for a quick demo and Dexter for instructions someone must follow.

I keep those jobs explicit. A recommendation can happen without spending credits. Producing a new recording uses your account, so approve the script and the amount of text before the tool runs.

Which Claude setup do I actually need?

Start with where you use Claude. Claude Desktop can connect to an MCP server. Claude Code can also work with local files and scripts, subject to its permissions. A chat with no connected generation tool can still write a script and hand you the voice’s share link; it cannot produce an ElevenLabs recording merely by naming the service.

For my catalog, the agent setup page gives the installation options and existing samples. Use it when the request is casting, such as a sixty-second product explainer or a chapter narrator. Keep the generation connection separate in your project notes so the next session knows what is installed and what is only a recommendation.

The repository skill installs with npx skills add bilbop1/prestige-voices. Its 1.1 workflow can offer to add a selected library voice to My Voices after your explicit approval, then optionally generate using your own key and credits. The separately submitted Claude directory bundle remains focused on casting and handoff. Installing that bundle does not add the repository’s account-action commands.

A useful first request is specific enough to produce an audition.

Pick three voices from Prestige Voices for a forty-second screen demonstration. Explain one weakness of each. Give me the existing preview links and a ninety-word draft. Do not generate audio yet.

That asks for something you can review. “Make this sound professional” leaves the assistant guessing about the format and gives you little basis for accepting its choice.

How does the hosted ElevenLabs connection work?

ElevenLabs documents a hosted MCP connector in Claude Desktop’s directory. Open Settings, then Connectors, locate ElevenLabs and connect. Complete the account authorization yourself after reviewing the requested permissions. The official page lists Text to Speech among the scoped capabilities, alongside agent-management operations.

After connecting, ask Claude to show the available speech tools and explain what the proposed operation will do. A connector used to configure a phone agent and a tool used to render an MP3 have different purposes. Confirm the tool available in your session supports the task before giving it a whole script.

Use the correct workspace and region. If your organization has isolated regional accounts, follow ElevenLabs’ regional connector instructions. A global account authorization won’t substitute for access to the regional workspace holding a project.

When would I use the local MCP server or a script?

ElevenLabs also maintains an official local MCP server. Its repository documents installation, local client configuration and output-folder settings. That route is useful when you need files written into an editing project or your environment already manages local credentials.

Keep API keys in the environment or approved local configuration. Don’t put one in the script, prompt, screenshot or repository. Ask the agent to confirm that a key is configured without displaying its value. If you need to debug authorization, share the error type after removing account details.

For a small repeatable workflow, a local script may be easier to inspect than a broad tool connection. The API quickstart shows the request and output handling. You can review exactly which voice, text and model the script sends. Both routes need a deliberate budget for retries.

Which voices would I audition for agent-written scripts?

These are three roles within my collection, chosen from the catalog’s delivery notes. They aren’t a ranking of the entire ElevenLabs library.

Jett is my first audition for a quick explanation. It measured about 180 words per minute in my clips. Its brisk delivery makes a duplicated point obvious, which is useful when an agent has overexplained a simple screen. Cut the duplicate before spending on another take.

Dexter gives a walkthrough an even read. It measured about 190 words per minute in my clips, quicker than Jett, so budget fewer spoken words when the viewer needs to click. Its evenness can make a launch pitch feel polite, so don’t reuse the same casting brief for a trailer.

Titan fits a serious technical or documentary passage, measured near 170 words per minute in my clips. A light weekly recap can sound too grave. If Jett is the better match, add Jett to My Voices and audition your own paragraph.

What should I approve before generation?

Ask Claude to show the final words, exact voice ID, model, character count and output path. Give a maximum number of takes. A request to fix a sentence should not accidentally regenerate a twenty-minute chapter.

For an audition, I’d use this fictional passage.

Open the project and choose the clip you want to replace. The original stays in the archive. Add the new recording on a separate track, line up its first word, and play the transition once without music. If the instruction finishes before the action on screen, keep the gap. The viewer needs that time to follow it. Save this as a new revision so the approved edit remains easy to find.

Then listen. Check whether “separate track” and “new revision” remain distinct, and whether the quiet beat is useful. The agent can report a saved file; your review settles whether it belongs in the edit.

How do I prevent an accidental publishing step?

Give the session a narrow finish line. Save the audition locally and stop. Treat uploading, emailing, posting or replacing production audio as another action. The proposed workflow should state those destinations plainly.

When the take is approved, save the script and voice recipe beside the audio. That gives the next Claude session enough context to revise one line without reinventing the casting or asking you to audition the same voices again.

Related guides

Other jobs where I use the same voices.

Nothing playing

Press play on any voice

↑ ↓ moveEnter openShift Enter playEsc