The Best MCP Servers for Audio Generation in 2026

Key Takeaways

• Best overall: Melius. One sign-in gives your assistant about 17 audio models from eight providers, including ElevenLabs speech, music and sound effects.

• Best for voice quality from the source: ElevenLabs. Its official hosted server signs in with OAuth and is listed in Claude's connector directory.

• Best for speech, music and effects in one model: Mage. Seed Audio makes dialogue, music beds and sound effects up to 120 seconds, alongside images and video.

• Best cheap commercial speech: Cartesia. Its Pro plan adds a commercial license and instant voice cloning for $5 a month.

• Check the license. Several free tiers here, including ElevenLabs' and Cartesia's, don't allow commercial use of what you generate.

What an audio MCP server does

The Model Context Protocol (MCP) lets an AI assistant such as Claude, ChatGPT or Cursor call outside tools. Connect an audio MCP server and you can ask your assistant to voice a script, score a clip or make a door-slam sound effect, and it calls the generation model for you and hands back a file or a download link.

That matters most when audio is one step in a bigger job. If Claude is already writing your video script, it can voice the narration in the same conversation instead of you copying text into a separate app. The servers below differ in three ways: how many kinds of audio they cover, whether they're hosted (sign in and go) or run locally on your machine, and how they bill.

How we picked

• Audio range. Speech, voice cloning, music and sound effects, and how many models sit behind each.

• Setup. Hosted servers with OAuth sign-in rank above local installs that need an API key and a Python or Node setup.

• Official support. We only included servers built by the company that runs the models, not community wrappers.

• Cost and rights. How generation is billed and whether you can use the output commercially. We checked each server's documentation and pricing on October 8, 2026.

The seven servers at a glance

1. Melius - the widest audio menu behind one connector

Melius, the AI-native operating system for creatives, runs a node-based canvas for image, video and audio, and its MCP server gives an assistant access to the same models. On the audio side its documentation lists ElevenLabs v3, Multilingual v2, Turbo v2.5, Sound Effects, Music, Scribe v2 transcription, Voice Changer and Voice Isolator, plus Seed Audio, Fish Audio S2.1 Pro, MiniMax Speech 2.8 HD and Music 3, Qwen3 TTS, Gemini 3.1 Flash TTS, xAI TTS and Sonilo sound effects.

Tools for listing voices, including cloned, designed and imported ones, mean your assistant can pick a voice by name, and Creator plans and up can import your own ElevenLabs voices. Because the audio sits next to Melius's image and video models, one conversation can produce a clip, its voiceover and its sound effects. The trade-off is distance from the vendors: Melius doesn't publish per-model length limits, and new vendor features arrive on its schedule rather than theirs.

Best for: Creators who want many voice, music and effects models without a separate account for each.

Audio it can make: Speech, voice changing, music, sound effects, transcription.

Setup: Hosted with OAuth at mcp.melius.com/mcp, an API-key endpoint, or a local CLI. Documented with Claude.ai, Claude Desktop, Claude Code, ChatGPT desktop and Codex.

Cost: Credits. A limited free trial, then Creator at $20 a month ($17 billed yearly). As of October 2026.

Link: melius.com

2. ElevenLabs - the official voice server

ElevenLabs now runs a hosted MCP server at api.elevenlabs.io/v1/mcp with OAuth sign-in and no API key, and it's listed in Claude's connector directory. It archived its older local server in August 2026 in favor of the hosted one, so skip any setup guide that tells you to install it with uvx.

The hosted server's documentation covers text to speech and ElevenLabs' agent tools today. ElevenLabs says the connector is expanding to its creative tools, including music, sound and video, but its own page still labels some of that as coming soon, so check what your connection shows before you plan a workflow around music. The free plan includes 10,000 credits a month but no commercial license, and free output you publish has to credit ElevenLabs.

Best for: Anyone who wants ElevenLabs voices straight from the source.

Audio it can make: Speech today; music and sound effects rolling out.

Setup: Hosted with OAuth. Install instructions cover Claude, Claude Code, ChatGPT, Cursor and others.

Cost: Free plan with 10,000 credits a month (non-commercial); Starter from $6 a month. As of October 2026.

Link: elevenlabs.io

3. Mage - speech, music and effects from one model

Mage's MCP server generates audio through the same tool it uses for images and video, with Seed Audio as its audio model. Seed Audio makes character dialogue and narration, music beds, sound effects and ambience, and audio matched to an image, in lengths from 5 to 120 seconds. You can pass up to three audio references or one image as a guide, and attach saved voices to Characters so the same character keeps the same voice.

A free cost-estimate tool lets your assistant check the price before it generates. Every MCP generation is paid in Gems, separate from the unlimited allowances in Mage's memberships, and there's no free plan to start from. Mage says commercial use is allowed.

Best for: Projects that need images, video and audio for the same characters.

Audio it can make: Speech, music, sound effects and ambience.

Setup: Hosted at mcp.mage.space/mcp with OAuth, or an API key in Claude Code and Cursor. Documented with Claude, ChatGPT, Claude Code, Cursor and Grok.

Cost: Gems, from the Basic plan at $10 a month with 500 Gems. As of October 2026.

Link: mage.space

4. fal - a model catalog with prices built in

fal's MCP server is a relay to its whole model catalog rather than an audio product. Your assistant can search models, read a model's inputs, check its price and run it, which makes it easy to try several text-to-speech or music models without leaving the chat.

There's no voice library or cloning workflow, and the license depends on the model you run, so check each one before using the output commercially.

Best for: Developers comparing audio models side by side.

Audio it can make: Depends on the model; text-to-speech is a searchable category.

Setup: Hosted with OAuth at mcp.fal.ai/mcp-relay. Documented with Claude, Claude Code, ChatGPT, Codex and Cursor.

Cost: The server is free; runs are billed at fal's API rates.

Link: fal.ai

5. Cartesia - cheap commercial speech with fine control

Cartesia's MCP server is speech only, but it's thorough: 18 tools cover text to speech, transcription, voice cloning, localizing a voice into another language and accent, and pronunciation dictionaries for names and jargon that other voices trip over.

The free plan gives 20,000 credits a month without a commercial license or cloning. The Pro plan at $5 a month adds both, which is the cheapest commercial entry on this list. There's no music or sound effects.

Best for: Narration and voiceover where pronunciation has to be right.

Audio it can make: Speech, cloning, transcription.

Setup: Hosted at mcp.cartesia.ai/mcp, or local with an API key.

Cost: Free plan (non-commercial); Pro $5 a month. As of October 2026.

Link: cartesia.ai

6. MiniMax - voice design from a description

MiniMax's open-source server exposes text to speech, voice cloning and voice design, which builds a new voice from a written description. It also includes image and video tools.

It runs locally, and your API host has to match the region your key was issued in, a common setup snag. MiniMax's API has music models, but the MCP server doesn't expose them, and MiniMax's own README now points users to its command-line tool instead.

Best for: Developers who want to design and clone voices from a local setup.

Audio it can make: Speech, cloning, voice design.

Setup: Local install with Python or Node and an API key. Documented with Claude Desktop, Cursor and Windsurf.

Cost: Billed through MiniMax's API.

Link: minimax.io

7. Hume - speech with acting directions

Hume's MCP server brings its Octave text-to-speech model into your assistant, with the option to give acting instructions such as "tired" or "sarcastic" alongside the text. You can design and save voices from a prompt, though cloning isn't part of the MCP.

It's the most expressive speech option here and the least mature server: a local install from a version 0.1 package, with speech only.

Best for: Character voices that need a specific delivery.

Audio it can make: Expressive speech, voice design.

Setup: Local install with npx and an API key. Documented with Claude Desktop, Cursor and Windsurf.

Cost: Billed through Hume's API.

Link: hume.ai

What we left off

• Suno and Stability AI. Neither has an official MCP server. Community wrappers exist, and some sign in with your account credentials, which we wouldn't recommend.

• Google's mcp-genmedia. It covers Gemini TTS and Lyria music, but Google labels it as not an officially supported product, and it needs a Vertex AI setup.

• Replicate. Its official server runs any model in its catalog, but it isn't built around audio.

Which one should you connect?

• You want the most models with one sign-in: Melius.

• You want ElevenLabs voices: the ElevenLabs hosted server.

• You're making images, video and audio for the same characters: Mage.

• You're testing models against each other: fal.

• You need cheap commercial narration: Cartesia.

• You want to design a voice from a description: MiniMax or Hume.