OpenTranscribe MCP

OpenTranscribe MCP

Own the recorder. Choose the intelligence.

OpenTranscribe is an open-source MCP server that routes audio to the speech-to-text model of your choice and returns one provider-independent transcript schema.

It is built around a simple idea: buying a great recorder should not lock you into one transcription subscription. Use Plaud, a phone, an open-source wearable, or any other audio source you are authorized to access, then choose Microsoft MAI, ElevenLabs Scribe, or Groq Whisper without changing the downstream workflow.

OpenTranscribe does not jailbreak hardware or bypass access controls. It works only with audio the operator is authorized to access.

How it works

OpenTranscribe keeps the integration boundary stable: the recorder supplies an audio file or HTTPS URL, the server selects or calls the requested speech-to-text provider, and downstream tools receive the same canonical transcript shape.

What v1 ships

• MCP Streamable HTTP at /mcp, with stateless operation and bearer or external OIDC authentication
• transcribe_audio, list_transcription_models, estimate_transcription_cost, get_transcript_chunk, and delete_transcript
• Microsoft MAI-Transcribe-2, ElevenLabs scribe-v2, and Groq Whisper adapters
• explicit capability negotiation, routing policies, retries, and observable fallbacks
• HTTPS URL passthrough when supported and a bounded streaming proxy otherwise
• SSRF controls, signed-URL redaction, zero content logging, and cost/resource limits
• disabled-by-default retention; optional memory or S3-compatible temporary result storage
• Docker, production-oriented Scaleway Serverless Containers Terraform, tests, and GitHub Actions CI

On the Linux desktop

An optional Debian package, OpenTranscribe Setup, bundles the Python runtime, the engine, and a native setup application. Install it, add a provider key, and connect a local MCP client — no Python, no uv, no configuration files. The engine then runs as one client-owned STDIO process that opens no network socket.

The hosted and command-line paths below are unaffected by it. See OpenTranscribe Setup on Linux; its release evidence lives in docs/desktop-acceptance.md, where every scenario starts at *Not tested*.

Ten-minute quickstart

Requirements: Python 3.12 and uv, or Docker.

git clone https://github.com/fbossiere/open-transcribe-mcp.git

cd open-transcribe-mcp

Create a private .env containing only the settings for your selected provider and a strong random MCP bearer token. For Microsoft:

OT_ENVIRONMENT=prod

OT_HOST=127.0.0.1 OT_DEFAULT_PROVIDER=microsoft OT_DEFAULT_MODEL=MAI-Transcribe-2 OT_MICROSOFT__ENDPOINT=https://YOUR-RESOURCE.cognitiveservices.azure.com OT_MICROSOFT__API_KEY=YOUR-KEY OT_SECURITY__AUTH_MODE=bearer OT_SECURITY__BEARER_TOKEN=YOUR-RANDOM-TOKEN

Use the complete Groq, ElevenLabs or multi-provider recipes for other services. Omit unused settings rather than leaving empty URLs or numbers from .env.example, which is a configuration inventory. Protect your file with chmod 600 .env.

Start the server:

uv sync

uv run open-transcribe-mcp

The MCP endpoint is http://localhost:8000/mcp; probes are available at /healthz and /readyz.

Run the included bilingual, two-voice synthetic recording through the connected provider:

uv run python examples/transcribe.py --token YOUR-RANDOM-TOKEN

The fixture and reference transcript are non-sensitive and redistributable. Pass another authorized public HTTPS audio URL as the first argument to use your own source.

Connect a FastMCP client:

import asyncio

from fastmcp import Client

async def main() -> None: async with Client("http://localhost:8000/mcp", auth="YOUR-RANDOM-TOKEN") as client: models = await client.call_tool("list_transcription_models", {}) print(models) result = await client.call_tool( "transcribe_audio", { "request": { "source": {"type": "url", "url": "https://example.org/authorized-audio.mp3"}, "provider": "auto", "routing_policy": "quality", } }, ) print(result)

asyncio.run(main())

Provider choice does not alter the response contract. Set provider and model to switch explicitly, or use auto with default, quality, cost, or latency routing.

diarization, timestamps, and transcript_style are unset above on purpose: an unset capability is not requested, so the request routes to any configured model and the response metadata reports what that model applied. Stating one makes it a requirement — "diarization": True excludes every model that cannot diarize rather than quietly returning a single-speaker transcript. See providers and capabilities.

Docker

Use the published image and the private environment file from the quickstart:

docker run --rm --env-file .env \

-e OT_ENVIRONMENT=prod -e OT_HOST=0.0.0.0 \ -p 127.0.0.1:8000:8000 \ ghcr.io/fbossiere/open-transcribe-mcp:1.2.3

This binds the host port locally; put an HTTPS proxy in front when hosting remotely. You can also build from source with docker build -t open-transcribe-mcp:1.2.3 .. Inject provider credentials at runtime. The image includes the s3 extra, but result storage stays disabled unless configured.

Deploy on Scaleway

The reference infra/scaleway module provisions a private image registry, a Serverless Container with scale-to-zero (0–3 instances), HTTPS ingress and health probes. It can optionally add a TTL-bound result bucket and runtime identity.

1. Create a dedicated Scaleway Project and a deployment API key with the documented IAM permissions. Supply SCW_ACCESS_KEY and SCW_SECRET_KEY to the deployment shell.
2. Copy infra/scaleway/terraform.tfvars.example to terraform.tfvars in that directory. Fill your project ID, provider keys, authentication settings and image_tag = "1.2.3". The template configures ElevenLabs and Groq; review the ElevenLabs retention choice before use.
3. Bootstrap the registry, copy the public release image into it, then review and apply the complete deployment. In infra/scaleway:

   terraform init

terraform apply -target=scaleway_registry_namespace.this IMAGE_REFERENCE="$(terraform output -raw image_reference)" REGISTRY_ENDPOINT="$(terraform output -raw registry_endpoint)" printf '%s' "$SCW_SECRET_KEY" | docker login "${REGISTRY_ENDPOINT%%/*}" \ --username nologin --password-stdin docker buildx imagetools create --tag "$IMAGE_REFERENCE" \ ghcr.io/fbossiere/open-transcribe-mcp:1.2.3 terraform plan -out=open-transcribe.tfplan terraform apply open-transcribe.tfplan terraform output -raw mcp_endpoint

1. Check /healthz, /readyz, MCP authentication and one short authorized transcription. The full Scaleway guide covers release digest verification, adoption of an existing deployment, OIDC setup, upgrades, rollback and result-store permissions.

Bootstrap applies are for new infrastructure. Import an existing registry, namespace and container into state first to preserve their IDs and endpoint. Keep real terraform.tfvars, state and saved plans private: state and plans can contain secrets even when output is redacted. Publishing a release does not automatically update your Scaleway deployment.

Configuration and service combinations

Choose the provider, MCP authentication and result storage independently. Provider keys stay on the server and are never accepted as tool arguments. See the full configuration guide for complete environment/Terraform examples, settings precedence, key permissions and troubleshooting.

OT_DEFAULT_PROVIDER and OT_DEFAULT_MODEL are preferences, not a provider allow-list. An explicit request can choose another configured model. Hosted requests permit transient fallback by default; use allow_fallback=false when audio must go to only one provider. To require French speaker turns, use language=fr, diarization=true and strict_capabilities=true; see the Scribe recipe. Groq cannot fulfill that requirement and is rejected rather than silently returning an undiarized transcript.

ElevenLabs requests default to zero retention, which needs an eligible Enterprise account. A standard account requires an explicit choice of OT_ELEVENLABS__ZERO_RETENTION=false after accepting provider-side retention. Disabling OpenTranscribe storage does not disable provider retention. See configuration and retention.

For ordinary HTTP, OT_* environment variables override a working-directory .env. Terraform