Whisper Transcribe

whisper-transcribe-mcp

MCP server for audio transcription using faster-whisper (local, free, offline) or OpenAI Whisper API (cloud, requires API key). Works with Claude Desktop and Claude Code on macOS, Windows, and Linux.

---

Prerequisites

macOS

Option A — uv (recommended):

brew install uv
# or
curl -LsSf https://astral.sh/uv/install.sh | sh

Option B — Python: Python 3.10+ is included in macOS 12.3+. You can also install it with brew install python.

---

Windows

Option A — uv (recommended):

winget install astral-sh.uv

Or download the installer from astral.sh/uv.

Option B — Python: Download Python 3.10+ from python.org. During installation, check "Add Python to PATH".

No need to install ffmpeg or any compiler — everything is bundled in the package.

---

Linux

Option A — uv (recommended):

curl -LsSf https://astral.sh/uv/install.sh | sh

Option B — Python:

# Debian/Ubuntu
sudo apt install python3.12 python3.12-venv

# Fedora
sudo dnf install python3.12

# Arch
sudo pacman -S python

No additional system dependencies required.

---

Installation

Option A — uvx (recommended, no permanent install)

uvx automatically downloads and installs the package in an isolated environment. Only requires uv to be installed.

# Local backend:
uvx "whisper-transcribe-mcp[local]"

# OpenAI backend:
uvx "whisper-transcribe-mcp[openai]"

# Both backends:
uvx "whisper-transcribe-mcp[all]"

Option B — pip

# Local backend:
pip install "whisper-transcribe-mcp[local]"

# OpenAI backend:
pip install "whisper-transcribe-mcp[openai]"

# Both backends:
pip install "whisper-transcribe-mcp[all]"

---

Use Cases

Case 1 — Local backend only (free, works offline)

Uses faster-whisper to transcribe locally. The model is downloaded from HuggingFace on first use (~74MB for base) and cached.

Install:

pip install "whisper-transcribe-mcp[local]"

Environment variables:

WHISPER_MODEL=base   # or tiny, small, medium, large-v3

---

Case 2 — OpenAI backend only (best accuracy, requires API key)

Uses OpenAI's whisper-1 model. Requires an API key and internet connection. No local model downloads.

Install:

pip install "whisper-transcribe-mcp[openai]"

Environment variables:

OPENAI_API_KEY=sk-...

---

Case 3 — Both backends (OpenAI if key present, local as fallback)

If OPENAI_API_KEY is set, OpenAI is used automatically. Otherwise falls back to local faster-whisper.

Install:

pip install "whisper-transcribe-mcp[all]"

---

Configuration

Claude Desktop

Config file location by operating system:

Add the entry inside "mcpServers":

Windows note: Claude Desktop runs in a restricted environment and may not have uvx in its PATH, and it may use a Python version (e.g. 3.14) for which ctranslate2 (a dependency of faster-whisper) does not yet have prebuilt wheels. Two fixes are required:

1. Use the full path to uvx.exe instead of just uvx. Run where.exe uvx in PowerShell to find it (usually C:\Users\<YourUser>\.local\bin\uvx.exe).
2. Force Python 3.12 via the --python 3.12 flag so that a compatible wheel is used.

Case 1 — Local:

macOS / Linux:

{
  "mcpServers": {
    "whisper-transcribe": {
      "command": "uvx",
      "args": ["whisper-transcribe-mcp[local]"],
      "env": {
        "WHISPER_MODEL": "base"
      }
    }
  }
}

Windows:

{
  "mcpServers": {
    "whisper-transcribe": {
      "command": "C:\\Users\\<YourUser>\\.local\\bin\\uvx.exe",
      "args": ["--python", "3.12", "whisper-transcribe-mcp[local]"],
      "env": {
        "WHISPER_MODEL": "base"
      }
    }
  }
}

Case 2 — OpenAI:

macOS / Linux:

{
  "mcpServers": {
    "whisper-transcribe": {
      "command": "uvx",
      "args": ["whisper-transcribe-mcp[openai]"],
      "env": {
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Windows:

{
  "mcpServers": {
    "whisper-transcribe": {
      "command": "C:\\Users\\<YourUser>\\.local\\bin\\uvx.exe",
      "args": ["--python", "3.12", "whisper-transcribe-mcp[openai]"],
      "env": {
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Case 3 — Both (OpenAI takes priority if key is set):

macOS / Linux:

{
  "mcpServers": {
    "whisper-transcribe": {
      "command": "uvx",
      "args": ["whisper-transcribe-mcp[all]"],
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "WHISPER_MODEL": "base"
      }
    }
  }
}

Windows:

{
  "mcpServers": {
    "whisper-transcribe": {
      "command": "C:\\Users\\<YourUser>\\.local\\bin\\uvx.exe",
      "args": ["--python", "3.12", "whisper-transcribe-mcp[all]"],
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "WHISPER_MODEL": "base"
      }
    }
  }
}

Restart Claude Desktop after editing the file.

---

Claude Code

Works the same on macOS, Windows, and Linux. Requires uv installed.

Claude Code config file location:

The easiest way to add the server is via the Claude Code CLI, which updates the config file automatically:

# Case 1 — Local:
claude mcp add whisper-transcribe uvx -- "whisper-transcribe-mcp[local]"

# Case 2 — OpenAI:
claude mcp add whisper-transcribe uvx --env OPENAI_API_KEY=sk-... -- "whisper-transcribe-mcp[openai]"

# Case 3 — Both (OpenAI with local fallback):
claude mcp add whisper-transcribe uvx --env OPENAI_API_KEY=sk-... --env WHISPER_MODEL=base -- "whisper-transcribe-mcp[all]"

Windows + [all]: Add --python 3.12 before the package name to avoid ctranslate2 wheel issues. Edit ~/.claude.json directly and use "args": ["--python", "3.12", "whisper-transcribe-mcp[all]"].

To add it globally (available in all projects), use --scope user:

claude mcp add --scope user whisper-transcribe uvx -- "whisper-transcribe-mcp[local]"

Or edit ~/.claude.json directly and add inside "mcpServers":

{
  "mcpServers": {
    "whisper-transcribe": {
      "command": "uvx",
      "args": ["whisper-transcribe-mcp[local]"],
      "env": {
        "WHISPER_MODEL": "base"
      }
    }
  }
}

---

Environment Variables

Backend selection and fallback ([all] only)

When installed with [all], the backend is chosen at startup:

• OPENAI_API_KEY set → OpenAI is used. If the API call fails at runtime (network error, invalid key, quota exceeded), the server automatically falls back to local faster-whisper and includes a "fallback_reason" field in the response.
• OPENAI_API_KEY not set → local faster-whisper is used directly, no fallback attempted.

---

Security and Permissions

The server runs locally over stdio and only needs the access its tools imply:

Tool inputs are validated before use: model_size must be a known Whisper model name, language must be an ISO 639 code, and extension must be a known audio format. OpenAI errors are returned as a short summary (error type and HTTP status); the full message goes to stderr. Releases are published from GitHub Actions with PyPI Trusted Publishing and PEP 740 provenance attestations.

---

Available Tools

transcribe_file

Transcribes an audio file by path (mp3, wav, m4a, ogg, flac, webm, etc.).

Parameters:

• file_path (required): Absolute path to the audio file
• language (optional): Language code (es, en, fr, etc.). Auto-detected if not provided.
• model_size (optional): Local model size. Ignored with the OpenAI backend.
• post_process (optional, default false): If true, passes the transcription through GPT-4.1 to fix spelling, grammar, and punctuation. Requires the openai package ([openai] or [all]).
• post_process_prompt (optional): Custom system prompt for GPT post-processing. Use it to provide domain-specific context, proper nouns, or product names that Whisper may have misspelled. Falls back to a generic correction prompt if not provided.

Response (without post-processing):

{
  "text": "Full transcription...",
  "language": "en",
  "language_probability": 0.99,
  "segments": [
    { "start": 0.0, "end": 4.2, "text": "First segment..." }
  ],
  "backend": "local",
  "model": "base"
}

Response (with post_process: true):

{
  "text": "Corrected transcription...",
  "raw_text": "Original transcription from Whisper...",
  "post_process_model": "gpt-4.1",
  "language": "en",
  "language_probability": 0.99,
  "segments": [...],
  "backend": "local",
  "model": "base"
}

If post-processing fails, text retains the original transcription and a post_process_error field is added.

---

transcribe_base64

Transcribes audio provided as a base64-encoded string. Useful for programmatic integrations.

Parameters:

• audio_base64 (required): Base64-encoded audio data
• extension (optional, default mp3): File extension (mp3, wav, ogg, etc.)
• language (optional): Language code
• model_size (optional): Local model size
• post_process (optional, default false): Same as in transcribe_file.
• post_process_prompt (optional): Same as in transcribe_file.

---

list_models

Shows the active backend configuration, available local models, and the GPT model used for post-processing.

---

Local Model Sizes

Models are downloaded automatically from HuggingFace on first use and cached locally.

---

Troubleshooting

MCP not loading in Claude Desktop on Windows

Symptom: The server fails to start with a dependency resolution error like:

ctranslate2>=4.6.1 has no wheels with a matching platform tag (e.g., `win32`)
hint: You require CPython 3.14 (`cp314`), but we only found wheels for `ctranslate2` with: `cp39`, `cp310`, `cp311`, `cp312`, `cp313`

Cause: Two issues combined:

1. Claude Desktop does not include the user's local bin in its PATH, so uvx must be referenced by full path.
2. Claude Desktop's uvx may pick a Python version (e.g. 3.14) for which ctranslate2 — a native dependency of faster-whisper — does not yet have prebuilt wheels for Windows.

Fix: Use the full path to uvx.exe and force Python 3.12 explicitly:

"whisper-transcribe": {
  "command": "C:\\Users\\<YourUser>\\.local\\bin\\uvx.exe",
  "args": ["--python", "3.12", "whisper-transcribe-mcp[local]"],
  "env": { "WHISPER_MODEL": "base" }
}

To find your exact uvx.exe path, run in PowerShell:

where.exe uvx

Transcribing audio files in Claude Desktop

Symptom: Claude Desktop fails to transcribe an uploaded audio file. It may attempt to read the file as base64 and pass it to transcribe_base64, which then fails or hangs for files larger than ~50 KB.

Cause: Claude Desktop runs in a sandboxed Linux container. When you upload a file using the attachment button, it is stored at a path like /mnt/user-data/uploads/audio.mp3 — inside the container. The MCP server runs on your Windows machine and has no access to that container path. Claude's fallback of base64-encoding the file and passing it to transcribe_base64 fails in practice because even a small audio file produces hundreds of kilobytes of base64 text, which overflows the context window before the tool call can be made.

Fix: Do not use the attachment button to upload audio files. Instead, place the file anywhere on your Windows filesystem and reference its path directly in the message:

"Transcribe the file at C:\Users\YourUser\Downloads\audio.mp3"

The MCP server will read the file directly from Windows and send it to the transcription backend. This works for files of any size within the Whisper API limit (25 MB).

---

Development

This project uses uv for reproducible development environments:

uv sync --group dev
uv run ruff check .
uv run ruff format --check .
uv run pytest -q
uv run pre-commit install

See CONTRIBUTING.md for the contribution workflow.

Distribution and releases

The package is distributed through PyPI and described by server.json for the official MCP Registry. Version tags publish both destinations through GitHub OIDC, without long-lived publishing tokens. See docs/publishing.md for the release checklist and one-time repository setup.

The Registry entry describes the base PyPI package. Choose the [local], [openai], or [all] extra from the installation examples above so the transcription backend you need is installed.

---

License

MIT — see LICENSE