# OpenTranscribe MCP [Health: Active]

**Category:** 🎙️ Speech-to-Text  
**Repository:** https://github.com/fbossiere/open-transcribe-mcp  
**GitHub Stars:** 1  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/opentranscribe-mcp

## Description
Provider-independent speech-to-text with normalized transcripts over MCP.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "opentranscribe-mcp": {
    "command": "uvx",
    "args": ["open-transcribe-mcp"],
    "env": {
      "OT_ENVIRONMENT": "",
      "OT_HOST": "",
      "OT_DEFAULT_PROVIDER": "",
      "OT_DEFAULT_MODEL": "",
      "OT_SECURITY__AUTH_MODE": "",
      "OT_SECURITY__BEARER_TOKEN": ""
    }
  }
}
```

**Requires environment variables:** `OT_ENVIRONMENT`, `OT_HOST`, `OT_DEFAULT_PROVIDER`, `OT_DEFAULT_MODEL`, `OT_SECURITY__AUTH_MODE`, `OT_SECURITY__BEARER_TOKEN` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What OpenTranscribe MCP does

OpenTranscribe MCP provides a speech-to-text boundary between audio sources and downstream MCP clients. It accepts an audio file or authorized HTTPS URL, sends the audio to a configured transcription provider, and normalizes the result into one transcript shape. Supported adapters include Microsoft MAI-Transcribe-2, ElevenLabs Scribe v2, and Groq Whisper.

The design lets operators use audio from recorders, phones, wearables, or other authorized sources without tying later workflows to one transcription subscription. Provider selection can be explicit, or the server can choose using a routing policy for quality, cost, or latency. Requests may also require capabilities such as diarization, timestamps, or a transcript style; unsupported models are excluded rather than silently returning an unmet capability.

## How it works

The hosted service exposes MCP Streamable HTTP at `/mcp` and operates statelessly. Clients authenticate with a bearer token or an external OIDC arrangement. Health and readiness probes are available at `/healthz` and `/readyz`.

Audio can be passed through as an HTTPS URL when a provider supports that path. Otherwise, OpenTranscribe MCP uses a bounded streaming proxy. The server includes capability negotiation, routing policies, retries, and observable fallback behavior. Transcript responses include metadata describing the capabilities applied by the selected model.

Optional result retention is disabled by default. When enabled, temporary results can use memory or S3-compatible storage, with support for TTL-oriented deployment configurations.

## Setup and configuration

Run the application from source with Python 3.12 and `uv`, or use Docker. A typical source installation clones the repository, creates a private environment file, runs `uv sync`, and starts the process with `uv run open-transcribe-mcp`. The default local endpoint is `http://localhost:8000/mcp`.

Configuration supplies the selected provider, model, provider endpoint and credentials, plus MCP authentication settings. The example Microsoft configuration uses `OT_DEFAULT_PROVIDER=microsoft`, `OT_DEFAULT_MODEL=MAI-Transcribe-2`, a Microsoft endpoint and API key, bearer authentication, and a randomly generated bearer token. Groq, ElevenLabs, and multi-provider configurations use their corresponding provider settings. Unused settings should be omitted, and credential files should be protected from other users.

Docker images can receive configuration through an environment file and runtime environment variables. The image includes the S3 extra, but storage remains off until configured. A reference Terraform deployment targets Scaleway Serverless Containers with HTTPS ingress, health probes, and optional result storage.

On Linux, the optional OpenTranscribe Setup Debian package bundles the runtime and a native setup application. It runs the engine as a client-owned STDIO process and does not open a network socket.

## Tools and capabilities

The MCP tool surface includes:

- `transcribe_audio` for submitting audio and selecting providers, models, routing policies, and requested capabilities.
- `list_transcription_models` for discovering configured models and their capabilities.
- `estimate_transcription_cost` for estimating provider transcription costs.
- `get_transcript_chunk` for retrieving part of a stored transcript.
- `delete_transcript` for removing a stored transcript.

Operational controls include SSRF protections, signed-URL redaction, zero content logging, and cost and resource limits. Authentication, retention, and storage behavior can be configured for the deployment.

## Limitations and notes

Only audio the operator is authorized to access should be submitted. OpenTranscribe MCP does not bypass recorder access controls or defeat hardware protections. A requested capability can narrow provider selection; for example, requiring diarization excludes models that cannot provide it. Result storage is opt-in, and remote deployments should use an HTTPS proxy or equivalent secure ingress.

_Full upstream README: https://allmcps.com/mcp/opentranscribe-mcp/readme_