All writing

ElevenLabs MCP: connect and save a speaker-labeled timestamped transcript

Learn how to connect the ElevenLabs MCP to ChatGPT or Claude to generate and manage transcripts with speaker labels and timestamps.

Last updated

Menus and supported features can change. Compare the source-check date below with your current app version.

App · APP
Official documentation · not hands-on tested Sources checked:

Transcribe one recording with ElevenLabs MCP

This reader-requested task follows ElevenLabs MCP official documentation, checked on 2026-11-07. The exact first release date and deployed build are unverified; selection does not imply a recent launch. Documentation edit dates are not release dates. This is a documentation-based guide, not a hands-on test or a paid purchase.

The goal is to submit one short English recording that you have permission to use, save the text actually returned, and compare it with the original audio. This guide does not supply an audio file or a completed transcription: the reader must provide the recording. The request and manual saving method below are editorial suggestions, not confirmation that every product capability has been tested.

1. Prepare your account and client

ElevenLabs MCP is a hosted server connecting an ElevenLabs workspace to an AI assistant F1. The documentation describes OAuth sign-in and scoped workspace access without running a local server or copying an API key into the client F3.

Prepare your ElevenLabs account and chosen AI assistant, then open the official connection page. The announcement names supported clients including ChatGPT, Claude and Cursor F2. Account-specific availability and billing conditions have not been verified here. Before running a task, check the official pricing page, terms, and your account for any app fees and model usage charges. A successful connection does not guarantee a free task.

2. Connect ChatGPT or Claude

For ChatGPT, follow the official instructions to add ElevenLabs from the plugins directory and sign in with your ElevenLabs account. The official page describes a one-time OAuth connection that can be revoked later F6. Use the service’s sign-in flow rather than pasting passwords or API keys into the conversation.

In Claude, go to Customize → Connectors, add a custom connector named ElevenLabs, and enter https://api.elevenlabs.io/v1/mcp F7. For Claude Code, the documented alternative is to run claude mcp add --transport http elevenlabs https://api.elevenlabs.io/v1/mcp in your terminal, then use /mcp inside a session to sign in F8. This is distinct from adding an external MCP server to an ElevenLabs voice agent.

Sign in with the account for your environment, check the workspace and requested permissions, and complete OAuth F13. The ElevenLabs connector in Claude’s connector directory is described as a restricted version without image, video or music generation F5. Do not assume every connection exposes the same tools.

3. Inspect the actual tools and prepare the example input

Ask the connected assistant to inspect the actual transcription tool’s input schema, file transfer method and returned fields before making a request. If no attachment or transcription tool is exposed, stop here rather than inventing a capability or result. If the tool description does not name a model, record “MCP model unverified.” The web Speech to Text documentation names Scribe v2 as the website’s default model, but that does not establish the model used through MCP F12.

Example input

This is the shared editorial example for this series. Its duration is not an official recording requirement, and no official sample file is supplied. The reader must prepare a recording they have permission to make and send to the service. Check that the actual filename and contents match the description before using it.

Owned input: two-speakers.wav, a 20-second English recording made with permission. Speaker A says: The meeting starts at ten. Speaker B says: I will bring the notes. Request: Transcribe this file once, preserve the words, show speaker labels and available timestamps, and mark uncertain words instead of guessing.

4. Submit once and follow the original task

Once you have confirmed that the tool can accept the actual file, attach the recording once and send the request above once. The official announcement says that passing a recording to Scribe returns a transcript with speaker labels and timestamps F4. That describes a documented capability, not the result of running this example or a guarantee of a particular output format.

Expected result

Look for text corresponding to the original recording and whatever speaker and timing information is actually returned. Check the real response and tool description for word or sentence start/end fields, JSON structure and file export controls. This guide does not invent sample timestamps or a transcription result.

If the job is pending, follow the original task’s status. Do not resubmit the same file merely because output is delayed. If no status-checking function is available, record that limitation. If the result is missing or the task failed, do not move on to saving: record the error and incomplete status. Do not invent missing text or guess inaudible words.

5. Save the returned text and reopen it

If text has actually been returned and automatic export is unavailable, copy it into a normal text editor and save it as owned-transcript.txt in UTF-8. This is an editorial suggestion for manual file preparation, not a built-in MCP export feature. Preserve timing JSON in its original form only if it was actually returned.

Completion criteria

Reopen the saved file and compare its words and speaker changes with the original audio. If timestamps were returned, compare them with the playback positions; otherwise record timing information as unverified. If speaker or timing information is missing, do not mark every requested outcome as fulfilled. Flag uncertain words and sections requiring review, and keep the final file associated with its source audio. Accuracy, processing time and file creation have not been measured or tested for this guide.

6. Check connection problems and the web alternative

If authentication fails, compare the account environment, server URL, selected workspace and requested permissions with the official setup instructions F13. This is an editorial troubleshooting check, not a guarantee that signing in again will fix the problem. If the setup appears correct but file transfer or transcription tools remain unavailable, stop the MCP task and consider an available official route.

To compare the web alternative, use the same recording and review the Speech to Text flow: open Transcribe files and select the language or leave it on Detect F9. The web instructions say to open results by clicking the uploaded filename, click a word to play audio at that point, and download results through Export F10. Plain text, JSON, HTML, SRT and VTT are documented export options in the web Transcript Editor F11. Do not assume an MCP job automatically appears in web history or uses the same menus.

If you perform the comparison yourself, record speaker and timing information in the output, processing time, client compatibility, file transfer friction, automation capabilities and the intended audience’s needs using the same criteria. Time is unmeasured here, and neither route is declared faster or more accurate. First complete one request and inspect the actual result, then decide which workflow fits your needs.

ElevenLabs MCP documentation reading guide

Original BumbleV reading guide illustration; not a product screenshot or a result from our testing. Official release behind this illustration · CC BY 4.0 · 2026-11-06

Introduction

One-task guide

Documentation comparison

ElevenLabs MCP AI conceptual illustration — not a screenshot
AI-created conceptual illustration by BumbleV, not an application screenshot or execution result. No real screenshot with verified reuse terms was available. Instructions follow official documentation; expected outputs still require your own verification. Official release supporting the explanation · web-5427240d · CC BY 4.0 · 2026-11-07 ·

Sources & further reading