Deepgram MCP

Deepgram builds speech and audio APIs, and its MCP server exposes them to assistants and coding agents that speak the Model Context Protocol. Through it a client can transcribe prerecorded audio from a local file or a URL, generate speech from text, and run audio understanding passes over a recording, covering things like speaker diarization, topic detection, summarization and sentiment. One design choice sets it apart from most vendor servers: it fetches its tool list from Deepgram's API when a client connects instead of hardcoding it, so capabilities added on the platform side appear on the next connection without upgrading the package. It installs standalone and also ships bundled with Deepgram's command line tooling, which wraps the same server behind a subcommand, so teams already using the CLI do not need a second install. The fit is twofold. Developers building voice features get an agent that can run a real transcription against real audio before any integration code is written, which catches formatting and accuracy surprises early. Everyone else gets a way to work through a pile of recordings, calls, interviews, meetings or podcast episodes, conversationally rather than through a dashboard. Audio has to reach Deepgram to be processed, so treat it like any other API call that leaves the machine and check what a directory contains before pointing an agent at all of it.