Deepgram MCP Server

Official Deepgram MCP server giving AI editors access to speech-to-text, text-to-speech, and audio intelligence tools, with dynamic runtime tool discovery.

AI/ML by Deepgram API Key active

GitHub Docs

Overview

Deepgram's Model Context Protocol server lets AI coding assistants (Claude Code, Cursor, Windsurf, and other MCP-aware clients) talk directly to Deepgram's speech and audio intelligence APIs. It is published as the deepgram-mcp Python package and is also bundled inside the official Deepgram CLI (deepctl), where it is exposed via the dg mcp command. Capabilities span transcription (speech-to-text), speech synthesis (text-to-speech), audio intelligence/text analysis (Read), and account/project/model/usage management.

A notable design choice: the server fetches its tool list from Deepgram's API at runtime rather than hardcoding it in the package. New endpoints added to Deepgram appear in the client the next time it connects, without needing a package upgrade. The server supports both stdio (default) and SSE transports, runs locally, and authenticates with a Deepgram API key set via the DEEPGRAM_API_KEY environment variable or --api-key flag. When launched through the CLI, dg login handles auth automatically.

The repository is the official Deepgram repo at github.com/deepgram/mcp and is MIT licensed. Because tools are dynamically fetched, the README does not enumerate a fixed tool list, but the underlying CLI surface includes listen (transcribe), speak (TTS), read (text intelligence), models, projects, keys, members, usage, and requests.

Tools

Tool	Description
`listen / transcribe`	Transcribe audio from a URL or local file using Deepgram speech-to-text models (e.g. Nova).
`speak / text-to-speech`	Synthesize speech from text using Deepgram TTS (Aura) voices.
`read / text intelligence`	Run text analysis such as sentiment, topics, intents, and summarization on transcribed or arbitrary text.
`models`	List available Deepgram models for STT and TTS.
`projects`	List, inspect, and manage Deepgram projects on the authenticated account.
`keys / members`	Manage API keys and project members.
`usage / requests`	Query API usage, billing summaries, and request history for a project.
`dynamic tools`	Additional tools are fetched from the Deepgram API at runtime; new server-side tools appear without a package upgrade.

Setup Guide

Prerequisites

A Deepgram API key (free tier at console.deepgram.com)
Python (for the standalone deepgram-mcp package) or the Deepgram CLI

Option 1: Standalone package

pip install deepgram-mcp

Run it:

deepgram-mcp --api-key your_api_key_here

Option 2: Deepgram CLI (recommended)

pip install deepctl
# or: uv tool install deepctl  /  pipx install deepctl
dg login
dg mcp

Claude Code / Cursor / Windsurf config

Add to .mcp.json (Claude Code), .cursor/mcp.json (Cursor), or the equivalent Windsurf MCP config:

{
  "mcpServers": {
    "deepgram": {
      "command": "deepgram-mcp",
      "env": {
        "DEEPGRAM_API_KEY": "your_api_key_here"
      }
    }
  }
}

Or, using the CLI wrapper:

{
  "mcpServers": {
    "deepgram": {
      "type": "stdio",
      "command": "dg",
      "args": ["mcp"]
    }
  }
}

Quick add for Claude Code:

claude mcp add deepgram -- deepgram-mcp

Runtime flags

Flag	Default	Purpose
`--transport`	`stdio`	`stdio` or `sse`
`--port`	`8000`	SSE port
`--host`	`127.0.0.1`	SSE host
`--api-key`	env var	Deepgram API key
`--base-url`	`https://api.dx.deepgram.com`	API endpoint
`--debug`	`false`	Verbose logging

Use Cases

Transcribe a podcast or meeting recording from a URL inside Claude Code and have the agent summarize, tag speakers, and write follow-up actions.
Generate voiceover audio files from blog posts or release notes by asking the agent to call Deepgram TTS and save the output locally.
Build and debug a voice agent project: list available Nova/Aura models, scaffold SDK code, and run live transcription tests from the editor.
Audit Deepgram usage and billing from inside an AI assistant, listing projects, API keys, and recent request volume.
Run sentiment, topic, and summarization analysis (Deepgram Read) over transcripts or arbitrary text directly in the agent loop.

Example Prompts

"Transcribe https://example.com/interview.mp3 with diarization and give me a speaker-labeled summary."
"Use Deepgram to generate an MP3 of this changelog using the Aura asteria-en voice and save it to ./release-notes.mp3."
"List all Deepgram STT models that support Spanish and pick the best one for phone-call audio."
"Run Deepgram Read on this support transcript and extract sentiment, topics, and a 3-sentence summary."
"Show my Deepgram API usage and request count for the last 30 days across all projects."

Pros

Official, MIT-licensed server maintained by Deepgram at github.com/deepgram/mcp.
Tool list is fetched dynamically from the Deepgram API, so new capabilities appear without upgrading the package.
Supports both stdio and SSE transports and ships inside the official dg CLI for one-command launch (dg mcp).
Auto-detects common AI agents (Claude Code, Aider, Codex) and switches to clean JSON output and stderr-only logs.

Limitations

The README does not enumerate a fixed tool schema, so discoverability depends on the runtime catalog returned by the Deepgram API.
Requires a Deepgram account and API key; transcription, TTS, and intelligence calls consume paid credits beyond the free tier.
Python-only distribution (pip install deepgram-mcp / deepctl); no native Node or Docker image is documented.

Alternatives

ElevenLabs MCP for TTS and voice cloning.
OpenAI Whisper based community MCP servers for local speech-to-text.
AssemblyAI MCP community implementations for transcription and audio intelligence.