Deepgram MCP Server

The Deepgram MCP Server brings Deepgram's voice AI into your coding assistant. Deepgram specialises in speech-to-text, text-to-speech and voice agents, meaning it turns speech into text, text into speech and powers AI agents that talk. MCP, the Model Context Protocol, is an open standard that lets assistants such as Claude Code or Cursor work with outside services. Deepgram builds the server itself. It ships inside dg, the official command-line tool, and you start it with the command dg mcp. The repository is on GitHub under the MIT licence.

What is the Deepgram MCP Server good for?

The server is aimed mainly at people who build voice features into their own applications. If you are working on transcription for support calls, captions for videos or a phone agent, you would otherwise keep jumping between editor, console and API reference. With the server, Deepgram says the assistant reaches the developer API directly and works with your account. The docs server rounds this out: the assistant can look up parameters, models and examples there instead of guessing from outdated training data. For an API with many models and options, that saves a good number of failed attempts.

Two servers for two jobs

Deepgram offers two separate routes, and the difference matters. The first is dg mcp. This server runs locally, by default over stdio, which means your editor starts it as its own process. It can also run over SSE, an HTTP-based transport, on a port you choose. According to the repository, it is technically a proxy that accepts requests from the editor and passes them to the Deepgram developer API. The second route is a hosted server for the documentation at https://api.dx.deepgram.com/kapa/mcp. It gives the assistant access to all of the Deepgram docs, and the setup page does not mention any sign-in for it. So if you only want to look things up, you can skip the CLI altogether.

Which tools does the server offer?

Deepgram does not publish a fixed tool list. According to the documentation, the remote service manages the list, so you see the current tools and their schemas in your MCP client. The upside is that new features can appear without a CLI update. The downside is that the scope may change. According to its website, the CLI itself covers more than 25 commands, such as dg listen for transcription, dg speak for voice output and dg read for text analysis. The CLI website says all commands become available through MCP, whereas the documentation and repository describe the proxy to the developer API. So once you are connected, check which tools your client actually shows.

Skills as an add-on

Deepgram also maintains the deepgram/skills repository, with reference files that coding tools load when needed. It contains an API reference generated from the API specifications, a guide through the documentation, working sample apps and instructions for setting up MCP. In Claude Code, you can install the skills as a plugin. That may be all you need for plain coding against the API, whereas the MCP server works directly with your account.

Costs and privacy

The CLI and the server itself are free. Use of the Deepgram API, however, runs through your account and is billed by consumption. According to the pricing page, there is free starting credit with no credit card required, followed by pay-as-you-go, prepaid Growth plans and Enterprise plans. Privacy deserves a careful look. According to Deepgram, requests take part in its Model Improvement Program by default. Audio, text and transcripts are then retained and used to improve its models. Setting the parameter mip_opt_out=true prevents this on a per-request basis. The setup documentation does not say whether the MCP tools set this parameter.

Who should use the Deepgram MCP Server?

It suits teams that already work with Deepgram or are building a voice application and want their assistant wired directly into the API and docs. For an occasional first look, the docs server is often enough, since its setup page asks for neither a sign-in nor the CLI. The server is a weaker fit if you only want to transcribe the odd file, because the CLI alone handles that. With confidential recordings, settle the retention rules first and start with test data.

The Deepgram dg CLI (Python 3.10 or newer) and a Deepgram account with an API key, plus an MCP client such as Claude Code or Cursor. For the docs server alone, a client with HTTP transport is enough.

With the Deepgram MCP Server, your AI assistant reaches the Deepgram developer API. A second, hosted server gives it the documentation. This guide covers the setup in Claude Code and Cursor and points out what to check on privacy.

Requirements

You need Python 3.10 or newer, a Deepgram account and an API key. Deepgram says you can get the key for free in its console. Keep in mind that API usage is billed by consumption. According to the pricing page, there is free starting credit with no credit card required.

Step 1: Install the CLI

The simplest route is pipx install deepctl or uv tool install deepctl. On macOS and Linux, brew install deepgram/tap/deepgram works too, and Homebrew also installs ffmpeg and portaudio. Deepgram offers an install script at deepgram.com/install.sh as well. It is a good habit to read a script like that before you run it.

Step 2: Sign in

Run dg login and follow the prompts. The CLI stores the key locally. Use dg whoami to check your sign-in status. In CI environments, set the DEEPGRAM_API_KEY environment variable instead.

Step 3: Connect Claude Code or Cursor

In Claude Code, run claude mcp add --scope user deepgram -- dg mcp. The server is then available across all your projects. In Cursor, open .cursor/mcp.json and add a deepgram entry under mcpServers with type stdio, command dg and args ["mcp"]. Restart the client afterwards. In Claude Code, /mcp shows whether the connection works and which tools are available.

Step 4: Add the docs server

The hosted docs server helps the assistant look things up. In Claude Code, run claude mcp add deepgram-docs --scope project --transport http https://api.dx.deepgram.com/kapa/mcp inside your project. In Cursor, add a deepgram-docs entry to the same configuration file with type http and this address. You do not need the CLI for this.

Step 5: First tasks

Start with questions about the documentation, for example: Which model suits live transcription in German? Which parameters turn on speaker detection? Then check the available tools in your client. Deepgram manages that list, so it may change. Try API calls with harmless sample data first, in a separate project.

Work safely

According to Deepgram, API requests take part in its Model Improvement Program by default. Audio, text and transcripts are then retained and used to improve its models. The parameter mip_opt_out=true prevents this per request, and the opt-out can also be enforced across an account on request. The setup documentation does not say whether the MCP tools set this parameter. You can check it in the console under Usage > Logs, in the mip_opt_out field. Have paid calls confirmed in your client and give the agent its own key.

claude mcp add --scope user deepgram -- dg mcp

A Deepgram API key. Locally, dg mcp automatically uses the credentials saved with dg login. In CI environments, set the DEEPGRAM_API_KEY environment variable instead. The setup page for the docs server does not mention any sign-in.

The server works with the stored API key and therefore with its permissions in the Deepgram project. The remote service decides which tools are available. Where possible, give the agent its own key and use a separate project for testing.

Requests pass through the local proxy to the Deepgram developer API. According to Deepgram, data from API requests is retained for the Model Improvement Program by default, and mip_opt_out=true prevents this per request. The endpoint determines the region, and for the EU that is api.eu.deepgram.com. Metadata is available for 90 days and, according to Deepgram, contains no content. The setup documentation does not say whether the MCP tools use the opt-out or the EU endpoint.

Medium. The agent works with your API key and, depending on the tool, can trigger paid usage. Because the service manages the tool list, the scope can change without a CLI update, so check it in your client regularly and have calls confirmed. SSE mode binds to 127.0.0.1 by default and should not be exposed to a network unprotected. Review the curl install command before running it, or use pipx or Homebrew instead.