How to Use Scribe v3 in Windsurf
Scribe v3 turns recorded meetings, interviews, and videos into transcripts with speaker labels and word timestamps. Compared with Scribe v2, it improves transcription accuracy and speaker labeling.
Prompts to try with Scribe v3 in Windsurf
Copy an example into Windsurf, or ask in your own words.
Use Scribe v3 to transcribe recordings/product-review.mp4 with speaker labels. Compare the requested changes with the project files and draft an issue for each request, including its timestamp.
Use Scribe v3 to transcribe recordings/demo.wav in English. Compare the spoken walkthrough with README.md and propose documentation corrections with supporting timestamps.
Use Scribe v3 to transcribe recordings/customer-interview.mp3 with two expected speakers. Draft product feedback notes and keep customer quotes separate from your interpretation.
Use Scribe v3 to transcribe recordings/tutorial.mp4. Use the word timestamps to draft captions and save them as captions/tutorial.srt.
Use Scribe v3 to transcribe recordings/architecture-review.wav with speaker labels. Compare the discussion with docs/architecture.md and draft updates supported by transcript timestamps.
How to set up Scribe v3 in Windsurf
Pick the way that suits you. Each takes about two minutes.
Use Spiral's MCP server
- Add the server
Open the Customize page from the Agents window and click MCPs. Under This computer, open the config file, or edit it directly: ~/.config/devin/mcp_config.json on macOS and Linux, %APPDATA%\devin\mcp_config.json on Windows. Add this:
{ "mcpServers": { "spiral": { "url": "https://api.spiral.ai/mcp" } } } - Sign in
Click Authenticate on the Spiral server in Customize › MCPs, or let the agent call Spiral once: your browser opens. Sign in to Spiral, pick the organization and click Authorize. No API key.
devin mcp login spiral
- Check it
Spiral is listed under This computer in Customize › MCPs. From the terminal:
devin mcp list
Use the Spiral skill
- Send Windsurf this message
Your agent reads Spiral's skill and sets itself up.
set up https://spiral.ai/SKILL.md
- Sign in or paste a key
Approve Spiral in the browser, or give the agent a key from the console.
Use the Spiral CLI
- Install and sign in
Installs the Spiral CLI, signs you in and adds its skill to your coding agents.
npx -y spiral-cli@latest init
Call the Spiral API
- Create an API key
In the console, under API keys.
- Run Scribe v3
One request; the result comes back by polling or a webhook.
curl https://api.spiral.ai/v1/runs \ -H "Authorization: Bearer $SPIRAL_API_KEY" \ -H "Content-Type: application/json" \ -H "Idempotency-Key: $(uuidgen)" \ -d '{ "provider": "elevenlabs", "model": "scribe_v3", "input": { "file": "<file>" } }'
Then paste one of the prompts above: Windsurf finds Scribe v3 on Spiral, shows the price and runs it.
Scribe v3 providers and prices
Spiral runs Scribe v3 through ElevenLabs. Open a card for today's price, with every fee included.
Settings you can ask for
Scribe v3 works from your file and source URL. Ask Windsurf for any of these in your prompt.
| Setting | Allowed values | Default | Description |
|---|---|---|---|
| Seed | 0 to 2147483647 | If specified, our system will make a best effort to sample deterministically, such that repeated requests with the same seed and parameters should return the same result. Determinism is not guaranteed. Must be an integer between 0 and 2147483647. | |
| Diarize | on or off | off | Whether to annotate which speaker is currently talking in the uploaded file. |
| File format | pcm_s16le_16 or other | other | The format of input audio. Options are 'pcm_s16le_16' or 'other' For |
| No verbatim | on or off | off | If true, the transcription will not have any filler words, false starts and non-speech sounds. Only supported with scribe_v2 model. |
| Temperature | 0 to 2 | Controls the randomness of the transcription output. Accepts values between 0.0 and 2.0, where higher values result in more diverse and less deterministic results. If omitted, we will use a temperature based on the model you selected which is usually 0. | |
| Number of speakers | 1 to 32 | The maximum amount of speakers talking in the uploaded file. Can help with predicting who speaks when. The maximum amount of speakers that can be predicted is 32. Defaults to null, in this case the amount of speakers is set to the maximum value the model supports. | |
| Tag audio events | on or off | on | Whether to tag audio events like (laughter), (footsteps), etc. in the transcription. |
| Use multi channel | on or off | off | Whether the audio file contains multiple channels where each channel contains a single speaker. When enabled, each channel is transcribed independently. By default a separate transcript is returned per channel; set multichannel_output_style='combined' to instead receive a single transcript with all channels merged and sorted by time. Each word in the response includes a 'channel_index' field indicating which channel it was spoken on. A maximum of 5 channels is supported. Each channel is billed independently at the full audio duration, so cost scales linearly with the number of channels. |
| Detect speaker roles | on or off | off | Whether to detect speaker roles (agent vs customer). Requires diarize=true. Cannot be used with use_multi_channel=true. When enabled, speaker_id values will be 'agent' and 'customer' instead of 'speaker_0', 'speaker_1', etc. Usage incurs an additional 10% surcharge on base transcription cost. |
| Diarization threshold | 0.1 to 0.4 | Diarization threshold to apply during speaker diarization. A higher value means there will be a lower chance of one speaker being diarized as two different speakers but also a higher chance of two different speakers being diarized as one speaker (less total speakers predicted). A low value means there will be a higher chance of one speaker being diarized as two different speakers but also a lower chance of two different speakers being diarized as one speaker (more total speakers predicted). Can only be set when diarize=True and num_speakers=None. Defaults to None, in which case we will choose a threshold based on the model_id (0.22 usually). | |
| Timestamps granularity | none, word, or character | word | The granularity of the timestamps in the transcription. 'word' provides word-level timestamps and 'character' provides character-level timestamps per word. |
| Multichannel output style | separate or combined | separate | Controls the response shape when use_multi_channel is enabled. 'separate' (default) returns one transcript per channel under 'transcripts'. 'combined' merges all channels into a single transcript whose words are sorted by start time, each carrying a 'channel_index' - matching the single-channel response shape. 'combined' requires timestamps (timestamps_granularity must not be 'none') and does not support entity detection or redaction. |
Every API field of Scribe v3 is on its tool page: open its card above.
What can you do with Scribe v3 in Windsurf?
Turn recorded product discussions into timestamped evidence for implementation work. A coding agent can compare Scribe v3 transcripts with project files and draft issues or documentation from what was said.
Meeting decisions with supporting timestamps
Transcribe a recorded meeting with Scribe v3, then have your agent extract decisions and link each one to the relevant passage.
Interview quote selection
Use Scribe v3 to separate interviewer and guest speech. Ask your agent to collect candidate quotes with timestamps for review against the recording.
Podcast editing notes
Transcribe an episode with Scribe v3, then have your agent identify topic changes and passages to review during editing.
Video caption drafts
Use Scribe v3 word timestamps as the basis for caption timing. Have your agent divide the transcript into readable caption lines.
Lecture study notes
Transcribe a recorded lecture with Scribe v3. Ask your agent to organize the transcript into topics, definitions, and questions for review.
Support call review
Use Scribe v3 speaker labels to distinguish turns in recorded calls. Have your agent extract the reported problem, proposed resolution, and supporting timestamps.
Searchable voice notes
Turn recorded voice notes into text with Scribe v3, then ask your agent to add concise titles and topic labels.
Sound cues for accessible transcripts
Keep Scribe v3 audio event tags when laughter, applause, or music helps explain a recording. Have your agent preserve those cues in the transcript draft.
More on Scribe v3 and Windsurf
Overview
Scribe v3 in other agents
Questions about Scribe v3 in Windsurf
Can Windsurf use Scribe v3?
Yes. Add Spiral to Windsurf as an MCP server, and Windsurf can find and run Scribe v3 like any built-in tool.
How much does Scribe v3 cost in Windsurf?
$0.004335 per run at the default inputs for 60 seconds of input. You pay from your Spiral balance as you go, with no subscription, and a failed run is not charged. Your Windsurf plan is separate.
Do I need an ElevenLabs account to use Scribe v3 in Windsurf?
No. Spiral runs Scribe v3 on its own ElevenLabs account. You sign in to Spiral once and never paste a key.
How long does it take to set up Scribe v3 in Windsurf?
About two minutes: add the server, then sign in, then check it.
Does Windsurf wait while Scribe v3 runs?
Scribe v3 runs in the background. Windsurf starts the run, checks on it and gives you the json when it's ready.
What is Scribe v3 best used for?
Use Scribe v3 to transcribe recorded conversations, including meetings and interviews. Speaker labels and word timestamps help an agent turn the transcript into notes or locate passages in the recording.
How does Scribe v3 improve on Scribe v2?
ElevenLabs reports 27% lower word error rate across 102 languages on FLEURS, plus improved speaker labeling. Treat the benchmark as a comparison, then evaluate recordings that resemble your actual work.
Does Scribe v3 translate speech into English?
Scribe v3 transcribes speech in the language spoken. Ask your agent to translate the finished transcript as a separate step.
Can Scribe v3 distinguish several speakers?
Yes, Scribe v3 supports speaker labels for up to 32 speakers. Supplying the expected speaker count can help with predicting who speaks when.
How long can a recording be?
Scribe supports recordings up to 15 hours long. Split longer recordings into shorter sections before transcription.
Start using Scribe v3 in Windsurf
One Spiral account for every tool your agent needs. No subscription; you pay per run.
Paste this into Windsurf:
set up https://spiral.ai/SKILL.md


