Voxtract
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Voxtract transcribe the latest meeting recording in my inbox"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Voxtract
Extract structured knowledge from voice recordings.
Voxtract is an MCP server that transcribes audio files using Mistral's Voxtral model and lets your LLM agent handle the post-processing — cleaning, restructuring, summarizing — all within your existing setup. No extra AI agents, no extra costs beyond Mistral transcription at $0.003/minute.
Why Voxtract?
Dedicated voice-to-text products charge $80–180 for hardware plus monthly subscriptions for AI features. Voxtract takes a different approach:
Mistral handles transcription — $0.003/min (~$0.18/hour). Speaker diarization, 13 languages, context biasing for domain terms.
Your LLM handles intelligence — The agent you already use (Claude, GPT, etc.) does the post-processing. No extra API costs. Cleaning, restructuring, meeting notes, idea extraction — whatever you need.
MCP glues them together — Voxtract runs as an MCP server inside any MCP-compatible client (Claude Code, Claude Desktop, Cursor, Windsurf, or any agent framework supporting MCP). The LLM calls the tools directly, reads the transcript, and processes it in the same conversation.
The result: you record on your phone, drop the file in any project directory, and get organized markdown knowledge — for essentially free on top of what you already pay.
Related MCP server: vibevoice-asr
How it works
You speak → Audio file → Voxtract transcribes → LLM cleans & structures → Organized knowledgeVoxtract uses a session-based workflow:
Drop audio files in a working directory
Voxtract transcribes them via Mistral and stages the results
The LLM reads the transcript, picks a descriptive name, and finalizes into a clean directory
Optionally, the LLM post-processes the transcript (clean up, restructure, extract ideas)
The output is a named directory with your audio and markdown:
your-project/
quarterly-review-action-items/
recording.mpeg
transcript.md
transcript_clean.mdMultiple audio files that belong to the same topic get merged into a single transcript automatically.
Pluggable post-processing
The real power is in what happens after transcription. Voxtract ships with a clean_transcript MCP prompt, but the architecture is designed to be forked and customized.
The prompt templates live in a single file — src/mistral_voice_mcp/prompts.py — and each one is just a function that returns a message list. Want to turn Voxtract into a meeting assistant? A lecture note-taker? A medical dictation tool? Write your own prompt function, register it in server.py, done. Your LLM handles the rest.
# src/mistral_voice_mcp/prompts.py — add your own
def meeting_notes_messages(transcript: str) -> list[dict]:
return [{"role": "user", "content": f"Extract action items from:\n{transcript}"}]# src/mistral_voice_mcp/server.py — register it
@server.prompt(name="meeting_notes", description="Extract action items and decisions")
def meeting_notes(transcript: str) -> list[dict]:
return prompts.meeting_notes_messages(transcript)Some ideas for what you could build:
Meeting assistant → Extract decisions, action items, attendees
Lecture notes → Structured study material with key concepts
Lab notebook → Extract methods, observations, technical terms
Interview processor → Q&A format with key quotes highlighted
Brainstorm organizer → Turn scattered spoken ideas into coherent proposals
Since the LLM you already use is the post-processing engine, each new use case costs you zero extra — just a new prompt function.
Tools & Prompts
Tools
Tool | Description |
| Set the work directory. Creates |
| Show current workdir path and status counts |
| Set transcription language (e.g. |
| Show current language and timestamp status |
| Set domain-specific terms for transcription accuracy (max 100) |
| Show configured bias terms |
| Clear all bias terms |
| List audio files waiting in |
| Stage inbox files into a transcription session |
| List active staging sessions and their status |
| Transcribe all audio files in a session (with diarization) |
| Read merged transcript from staging before naming |
| Finalize session into a named directory with |
| List all finalized recording directories |
| Read transcript (raw or clean) from a recording |
| Save post-processed text as |
Prompts
Prompt | Description |
| Remove filler words, fix grammar and logic flow, restructure for clarity while preserving meaning and speaker headings |
Use it everywhere
The idea is simple: add Voxtract to any repository where you work. Got an idea while walking? Record it, drop the file, and get structured documentation — right next to your code, notes, or research.
Installation
No cloning, no setup. Requires uv and a Mistral API key (pay-per-use, no subscription).
# Available in all your projects
claude mcp add --scope user --env MISTRAL_API_KEY="your-key-here" voxtract -- uvx --from git+https://github.com/hitesit/voxtract voxtractOne command. Done. Use --scope project instead to limit it to the current project only.
Add this to your MCP config (.mcp.json for Claude Code, claude_desktop_config.json for Claude Desktop, or equivalent for Cursor, Windsurf, etc.):
{
"mcpServers": {
"voxtract": {
"type": "stdio",
"command": "uvx",
"args": ["--from", "git+https://github.com/hitesit/voxtract", "voxtract"],
"env": {
"MISTRAL_API_KEY": "your-key-here"
}
}
}
}uvx downloads, installs, and runs Voxtract in an isolated environment automatically.
Quick start
Once the MCP server is connected to your agent:
You: Set the work directory to /path/to/my/project
I dropped a voice memo in the inbox, transcribe it in Italian and clean it upThe agent will:
Set the workdir, list the inbox
Create a staging session, transcribe the audio
Read the transcript, suggest a directory name
Finalize into a named folder with
transcript.mdClean and restructure into
transcript_clean.md
All in one conversation, using tools you can see and control.
Supported formats
.flac, .m4a, .mp3, .mpeg, .ogg, .wav — up to 1 GB, up to 3 hours per file.
Supported languages
English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, Dutch.
Cost
Mistral transcription: $0.003 per minute of audio. A 1-hour meeting costs ~$0.18. Post-processing is handled by the LLM you already use — no additional cost.
License
MIT
Available Tools
16 toolsmistral_clear_context_biasADestructiveIdempotent
Clear all context bias terms.
Inputs: no named parameters. Returns: result
mistral_create_sessionA
Create a staging session from audio files in inbox/.
Copies the specified files (or all inbox files) into a staging session for transcription.
Inputs: filenames. Returns: result
mistral_finalizeA
Finalize a staging session into a named recording directory.
Creates a directory under workdir with the given name (slugified), moves audio files into it, writes the merged transcript.md, and cleans up the staging session and inbox originals.
Inputs: name (required), session_id (required). Returns: result
mistral_get_context_biasARead-onlyIdempotent
Show currently configured context bias terms.
Inputs: no named parameters. Returns: result
mistral_get_languageARead-onlyIdempotent
Show current transcription language and whether timestamps are active.
Inputs: no named parameters. Returns: result
mistral_get_workdirARead-onlyIdempotent
Show current work directory path and status.
Inputs: no named parameters. Returns: result
mistral_list_inboxARead-onlyIdempotent
List audio files in inbox/ waiting to be transcribed.
Inputs: no named parameters. Returns: result
mistral_list_recordingsARead-onlyIdempotent
List finalized recording directories.
Inputs: no named parameters. Returns: result
mistral_list_sessionsARead-onlyIdempotent
List active staging sessions.
Inputs: no named parameters. Returns: result
mistral_read_staging_transcriptARead-onlyIdempotent
Read the merged transcript from a staging session.
Use this to review the transcript before finalizing with a name.
Inputs: session_id (required). Returns: result
mistral_read_transcriptARead-onlyIdempotent
Read a transcript from a finalized recording.
Inputs: recording (required), clean. Returns: result
mistral_save_processedAIdempotent
Save a cleaned transcript into a finalized recording directory.
Writes the content as transcript_clean.md.
Inputs: content (required), recording (required). Returns: result
mistral_set_context_biasAIdempotent
Set context bias terms for better transcription accuracy.
Accepts comma-separated or newline-separated terms (max 100). If 'terms' is a file path relative to the work directory, reads terms from that file.
Inputs: terms (required). Returns: result
mistral_set_languageAIdempotent
Set the transcription language.
Default is English with timestamps enabled. Setting a non-English language automatically disables timestamps (Mistral API constraint: language and timestamp_granularities are mutually exclusive).
Inputs: language (required). Returns: result
mistral_set_workdirAIdempotent
Set the work directory for transcription. Creates inbox/ and .staging/ subdirectories if needed.
Inputs: path (required). Returns: result
mistral_transcribeA
Transcribe all audio files in a staging session.
Uses the session language set via mistral_set_language (default: English).
Inputs: session_id (required), diarize, timestamp_granularity. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 16 tools
Each tool targets a distinct stage of the transcription workflow: configuration (workdir, context bias, language), staging (inbox, sessions, transcription), and finalization (recordings, transcripts). Even the two transcript readers are clearly differentiated by staging vs finalized context.
Most tools follow a consistent mistral_verb_noun pattern (e.g., get_workdir, set_language, list_sessions). However, transcribe and finalize are bare verbs, and save_processed uses an adjective, which slightly breaks the pattern.
16 tools is just above the typical well-scoped range of 3-15, but each tool earns its place in the transcription pipeline. The count feels slightly heavy but not bloated.
The tool set covers the full lifecycle from inbox to finalized recording, with configuration options and staging review. Minor gaps exist (e.g., no explicit session deletion or transcript editing), but these can be worked around.
Maintenance
Related MCP Connectors
Transcribe audio & video to text for AI agents: 100+ languages, speaker labels, webhooks.
Audio for your agent: transcribe, speak, translate, summarise, plus sound effects and music.
AI transcription API: turn voice memos, meetings and audio URLs into text and SRT subtitles.
Transcribe audio and video and work with your transcript library from any AI app. 119 languages.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceThis service provides fast and reliable transcriptions for audio/video files and voice memos. It allows LLMs to interact with the text content of audio/video file.8MIT
- AlicenseAqualityDmaintenanceLocal speech-to-text transcription using Microsoft's VibeVoice-ASR model with speaker diarization, enabling audio transcription directly in AI tools like Claude Code, Cursor, and OpenCode.33MIT
- FlicenseNot gradedqualityDmaintenanceTranscribes audio/video files, generates summaries and structured knowledge items, and supports Notion integration and chat-based interaction. Works as a standalone CLI, Notion integration, or MCP server tool for agent ecosystems.-
- FlicenseNot gradedqualityDmaintenanceSelf-host a realtime voice-call companion for coding agents. Exposes an MCP endpoint that agents can poll as an alternate input stream.19-