Sanzaru (OpenAI audio MCP)

What it does

Community MCP server wrapping OpenAI's audio APIs: Whisper and GPT-4o transcription with enhancement templates, interactive 'chat with audio' analysis, format conversion and compression, and OpenAI TTS. A filmmaker-agent can transcribe dailies, interrogate them in place, then generate scratch narration without leaving the same credential. Its own blurb is now 'lightweight multimodal MCP server' rather than an audio one — recent releases have widened into gpt-image, ElevenLabs-voiced podcast synthesis, and a Sora video surface that no longer has a backend.

Connect

claude mcp add sanzaru -e OPENAI_API_KEY=YOUR_KEY -e SANZARU_MEDIA_PATH=/absolute/path/to/media -- uvx "sanzaru[all]"

Sample tools

Field notes

Successor to the widely-listed and now-deprecated arcaputo3/mcp-server-whisper. Still moving fast: PyPI sanzaru 0.12.2 (28 Sep 2026, three releases that day), MIT, Python >=3.10, but only 6 stars for a project of this scope, so it is one maintainer's momentum. Buy the audio, not the bundle: the video tools (create_video, remix_video and friends) target sora-2 and sora-2-pro, and OpenAI removed both models and the whole Videos API on 24 Sep 2026 with no replacement listed on its deprecations page — yet the README as of 29 Sep still presents them as current, so an agent that sees those tools will try them and fail. 0.12.1 did repair the audio side of the same churn: chat_with_audio had been broken on its defaults since OpenAI retired the gpt-4o audio-preview models, and now defaults to gpt-audio-1.5. 0.12.0 made ElevenLabs eleven_v4 the recommended podcast engine, and 0.11.1 added `wait_for` so agents block on long jobs instead of polling. The README now leads with `/plugin marketplace add TJC-LP/sanzaru` or `uv tool install sanzaru`; the uvx command above still works. Extras are video, audio, image, elevenlabs and databricks — pick `sanzaru[audio]` for transcription work and add `sanzaru[elevenlabs]` if you want ElevenLabs TTS instead of OpenAI's; `[all]` now installs a dead video module. No local moderation: content guardrails are OpenAI's. Every call bills your OpenAI account. For fully offline transcription use jwulff/whisper-mcp or SmartLittleApps/local-stt-mcp (whisper.cpp, Apple Silicon, diarization).

Also in Transcription & Captions

Deepgram MCPDeepgram Official

Official Deepgram server: Nova transcription, diarization, and audio intelligence.

transcription · diarizationfreemium

M MultiVoiceMultiVoice Official

Quoted dubbing, subtitles, transcription and multi-voice audiobooks from one remote MCP; a human approves each price.

transcription · captionsremote · paid

Pairs well with

Adobe for CreativityAdobe Official

Adobe's official hosted MCP — 50+ tools across Photoshop, Lightroom, Firefly, Premiere, Express and Stock.

image-editing · image-enhancementremote · freemium

Adobe Run-Workflow MCPAdobe Official

Adobe's second hosted MCP — compose and run Firefly Creative Production batch pipelines.

batch-processing · workflow-automationremote · paid

Clueso MCPClueso Official

Hosted OAuth MCP that builds and edits a real, re-openable timeline by conversation.

timeline-editing · clip-assemblyremote · freemium