Voice AI & Telephony10 servers Updated September 2026

Top 10 Voice AI & Telephony MCP Servers for AI Agents (2026)

Q: Can voice AI MCP servers place real telephone calls to customers?

Yes, servers like `twilio-mcp`, `vapi-mcp`, `retellai-mcp`, and `elevenlabs-mcp` include tools to initiate outbound PSTN telephone calls. Because these tools place live calls and incur carrier or platform charges immediately, teams should implement human-in-the-loop approval checkpoints before granting agents access to outbound dialing tools.

Q: Which voice and speech recognition MCP servers can run completely offline?

Both `kokoro-tts-mcp` for speech synthesis and `stt-mcp` or `mlx-whisper-mcp` for speech-to-text run completely on local hardware without sending data to cloud APIs. This ensures zero API costs and full data privacy for sensitive voice recordings.

The top 10 Model Context Protocol (MCP) servers for ElevenLabs, Twilio telephony, Cartesia streaming TTS, Vapi, Retell AI, and local Whisper speech recognition, with tool schemas and configs for Claude Desktop, Cursor, and OpenCode.

Each entry lists the tools a server registers, the command that installs it, the config block its host expects, and the limits it runs into: bot challenges, expired sessions, destructive calls that fire without approval.

What is in this list

Official protocol compliance. Evaluated for implementation of standard MCP transports over stdio and Streamable HTTP, supporting tool schemas, resources, and audio generation endpoints.
Tool schemas taken from source. Every callable tool name, argument signature, launch entrypoint, and environment variable is extracted directly from the repository source code.
Context and token discipline. Assessed on how responses handle large audio payloads, transcript streaming, and base64 payloads to prevent model context exhaustion.

Top 10 servers at a glance

Maintainer, stars, and transport for all 10. Select one to jump to its full entry.

#01

elevenlabs/elevenlabs-mcp

ElevenLabs (Official)

High-fidelity speech synthesis, voice cloning, audio isolation, and conversational voice agent orchestration

1.5kstdio

#02

@twilio-alpha/mcp

Twilio (Official Alpha)

Enterprise SMS messaging, voice call dispatching, and phone number administration via OpenAPI schemas

109stdio

#03

cartesia-mcp

Cartesia (Official)

Low-latency streaming speech synthesis, voice localization, accent control, and pronunciation management

14stdio / Streamable HTTP

#04

@vapi-ai/mcp-server

Vapi AI (Official)

Configuring conversational voice assistants, managing transcribers, and triggering outbound voice calls

57stdio

#05

retellai-mcp-server

abhaybabbar

Retell AI conversational agent management, dynamic LLM prompt configuration, and web call creation

39stdio

#06

notebooklm-mcp

Romain Peyrichou (roomi-fields)

Generating structured audio overviews, multi-speaker podcasts, and citation-backed summaries from source documents

170stdio / HTTP REST

#07

kokoro-mcp-server

Aparsoft

Completely offline, zero-cost text-to-speech synthesis and automated podcast generation using open weights

14stdio

#08

stt-mcp

sm18lr88

Local, private speech-to-text transcription with asynchronous job lifecycle management and CUDA safety policies

31stdio

#09

mlx-whisper-mcp

Kachi Odoemene

On-device speech-to-text transcription optimized specifically for Apple Silicon hardware

24stdio

#10

notebooklm-mcp (PleasePrompto)

PleasePrompto

Document research, citation extraction, and triggering Google NotebookLM deep research audio overviews

3.4kstdio

Speech Synthesis & Conversational AI

#01 elevenlabs/elevenlabs-mcp

ElevenLabs (Official) 1.5k stdio

Best for: High-fidelity speech synthesis, voice cloning, audio isolation, and conversational voice agent orchestration

The official ElevenLabs Model Context Protocol server. Built on FastMCP and the official Python SDK, it exposes 27 tools covering text-to-speech, speech-to-text with diarization, sound effect generation, conversational agent creation, and outbound call simulation.

Connects agents directly to ElevenLabs voice generation, sound effects, speech-to-text, and conversational AI endpoints.

With Webfuse. Pair ElevenLabs voice synthesis with Webfuse live browser sessions to narrate and execute web actions in real time.

Registered tools

text_to_speech(text, voice_name?, stability?: number, output_format?): Generate natural speech audio from text using ElevenLabs voice models.
speech_to_text(input_file_path, language_code?, diarize?: boolean): Transcribe audio files with speaker diarization.
create_agent(name, first_message, system_prompt, voice_id?): Deploy a conversational voice agent with system prompts and custom voices.
make_outbound_call(agent_id, agent_phone_number_id, to_number): Initiate an outbound phone call connecting a voice agent to a phone number.

Install

uvx elevenlabs-mcp

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "elevenlabs": {
      "command": "uvx",
      "args": ["elevenlabs-mcp"],
      "env": {
        "ELEVENLABS_API_KEY": "your-api-key",
        "ELEVENLABS_MCP_OUTPUT_MODE": "files"
      }
    }
  }
}

Limitations

Setting output mode to resources returns base64 audio in the prompt context, increasing token consumption.
Outbound telephony calls and voice cloning tools lack dry-run validation flags.

elevenlabs/elevenlabs-mcp PyPI

Works withClaude DesktopCursorOpenCodeVS Code Copilot

Telephony & SMS Communications

#02 @twilio-alpha/mcp

Twilio (Official Alpha) 109 stdio

Best for: Enterprise SMS messaging, voice call dispatching, and phone number administration via OpenAPI schemas

The official Twilio Alpha MCP server. Written in TypeScript, it translates Twilio OpenAPI definitions into tools that let agents send SMS messages, initiate programmable voice calls, manage phone numbers, and inspect call logs.

Translates Twilio OpenAPI specifications into callable MCP tools for programmable telephony.

With Webfuse. Trigger Webfuse authenticated browsing sessions when a customer requests support through Twilio SMS.

Registered tools

twilio_api_v2010--CreateMessage(To, From?, Body?): Send an outbound SMS message via the Twilio API.
twilio_api_v2010--CreateCall(To, From, Url?): Initiate an outbound programmable voice call connected to TwiML instructions.
twilio_api_v2010--ListMessage(To?, From?, PageSize?: number): Query sent and received SMS communication records.
twilio_api_v2010--ListCall(Status?, StartTime?, PageSize?: number): Retrieve voice call logs, durations, and completion statuses.

Install

npx -y @twilio-alpha/mcp

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "twilio": {
      "command": "npx",
      "args": [
        "-y",
        "@twilio-alpha/mcp",
        "-a",
        "ACXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX",
        "-k",
        "SKXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX",
        "-s",
        "your-api-secret",
        "--services",
        "twilio_api_v2010"
      ]
    }
  }
}

Limitations

Loading default services registers hundreds of tools that consume significant context tokens without filtering.
Destructive actions and billable SMS calls execute without separate approval prompts.

twilio-labs/mcp npm

Works withClaude DesktopCursorOpenCode

Real-Time Streaming TTS

#03 cartesia-mcp

Cartesia (Official) 14 stdio / Streamable HTTP

Best for: Low-latency streaming speech synthesis, voice localization, accent control, and pronunciation management

The official Cartesia Model Context Protocol server. Designed for real-time voice applications, it supports dual stdio and Streamable HTTP transports, returning downloadable audio URLs instead of embedding raw audio bytes directly into context.

Wraps Cartesia Sonic models with structured rate limiting and audio file references.

With Webfuse. Deliver low-latency voice responses in Webfuse co-browsing sessions using Cartesia streaming audio.

Registered tools

text_to_speech(transcript, voice_id, output_format: object, language?): Generate natural speech audio using Cartesia Sonic models.
localize_voice(voice_id, name, accent, language): Adapt existing voices to regional dialects and accents.
list_voices(limit?: number, gender?, language?): List available voices with metadata and pagination cursors.
speech_to_text(file_path, mode?, language?): Transcribe audio files with word-level timestamp precision.

Install

uvx cartesia-mcp

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "cartesia": {
      "command": "uvx",
      "args": ["cartesia-mcp"],
      "env": {
        "CARTESIA_API_KEY": "your-cartesia-api-key"
      }
    }
  }
}

Limitations

Requires Python 3.13 or newer, which may require environment adjustments on older host runtimes.
Voice cloning and change operations require audio files stored on the local MCP server filesystem.

cartesia-ai/cartesia-mcp PyPI

Works withClaude DesktopCursorOpenCode

MCP gives coding agents tool access.

Webfuse gives them live , authenticated browser execution.

Start Building FreeFree tier · no credit card

AI Agent

MCP

Find the claim form

Fill patient details

Submit claim

Task completed

Voice Assistant Platform

#04 @vapi-ai/mcp-server

Vapi AI (Official) 57 stdio

Best for: Configuring conversational voice assistants, managing transcribers, and triggering outbound voice calls

The official Vapi Model Context Protocol server. Running on Node with the Vapi server SDK, it provides tools to create voice assistants, configure transcribers and speech models, provision phone numbers, and place outbound telephone calls.

Manages Vapi voice agents, phone numbers, customer call dispatch, and dynamic assistant tools.

With Webfuse. Embed Vapi voice assistants inside Webfuse web sessions to let AI agents talk to customers while co-browsing.

Registered tools

create_assistant(name, instructions?, voice?: object, transcriber?: object): Create and configure a full conversational voice assistant.
create_call(assistantId?, phoneNumberId?, customer: object): Dispatch an outbound phone call connecting a customer to a voice assistant.
list_assistants: List active voice assistants configured in your Vapi organization.
get_call(callId): Retrieve call duration, status, cost, and execution details.

Install

npx -y @vapi-ai/mcp-server

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "vapi": {
      "command": "npx",
      "args": ["-y", "@vapi-ai/mcp-server"],
      "env": {
        "VAPI_TOKEN": "your-vapi-token"
      }
    }
  }
}

Limitations

List tools use fixed result caps without cursor pagination for large assistant directories.
Placing phone calls incurs telephony costs immediately without a separate dry-run preview parameter.

VapiAI/mcp-server npm

Works withClaude DesktopCursorOpenCode

Conversational Telephony Agents

#05 retellai-mcp-server

abhaybabbar 39 stdio

Best for: Retell AI conversational agent management, dynamic LLM prompt configuration, and web call creation

A community TypeScript MCP server for Retell AI. It exposes 21 tools covering agent creation, dynamic variable injection, phone number provisioning, and initiating web-based and PSTN telephone calls.

Bridges Retell AI agent configurations, phone numbers, and call analytics to MCP clients.

With Webfuse. Connect Retell web calls into Webfuse live web spaces for voice-guided website onboarding.

Registered tools

create_phone_call(fromNumber, toNumber, overrideAgentId?): Place an outbound phone call using a configured Retell agent.
create_web_call(agentId, metadata?: object): Generate a web call token for browser-based voice conversations.
create_agent(llmWebsocketUrl, voiceId, agentName?): Register a new conversational voice agent with Retell.
list_calls(agentId?, limit?: number, offset?: number): Query past call records, durations, and call analysis metrics.

Install

npx -y @abhaybabbar/retellai-mcp-server

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "retell": {
      "command": "npx",
      "args": ["-y", "@abhaybabbar/retellai-mcp-server"],
      "env": {
        "RETELL_API_KEY": "your-retell-api-key"
      }
    }
  }
}

Limitations

Resource deletion tools are exposed unconditionally without separate verification prompts.
Does not implement automatic request retries or backoff when hitting API limits.

abhaybabbar/retellai-mcp-server npm

Works withClaude DesktopCursorOpenCode

Audio Overview & Podcast Generation

#06 notebooklm-mcp

Romain Peyrichou (roomi-fields) 170 stdio / HTTP REST

Best for: Generating structured audio overviews, multi-speaker podcasts, and citation-backed summaries from source documents

A feature-rich MCP server and REST service for Google NotebookLM. It connects research notes, document sources, and study aids to generate two-speaker podcast discussions, audio overviews, and structured study guides.

Drives Google NotebookLM workflows to convert research notebooks into multi-speaker conversational audio.

With Webfuse. Scrape reference documentation through Webfuse proxies and pass clean markdown into NotebookLM for audio podcast generation.

Registered tools

generate_content(content_type, custom_instructions?, language?): Generate conversational audio overviews, reports, or study aids from notebook sources.
ask_question(question, notebook_id?): Query notebook documents with grounded, citation-backed answers.
add_source(notebook_id, source_type, content): Add a document, web URL, or text note to a notebook.
list_notebooks: List accessible notebooks and active document collections.

Install

npx -y @roomi-fields/notebooklm-mcp

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "notebooklm": {
      "command": "npx",
      "args": ["-y", "@roomi-fields/notebooklm-mcp"],
      "env": {
        "NOTEBOOKLM_CONTENT_LANGUAGE": "en"
      }
    }
  }
}

Limitations

Relies on browser automation and undocumented Google endpoints, requiring dedicated account cookies.
Subject to daily account generation quotas for long audio overviews.

roomi-fields/notebooklm-mcp npm

Works withClaude DesktopCursorOpenCode

Open-Weights Local TTS

#07 kokoro-mcp-server

Aparsoft 14 stdio

Best for: Completely offline, zero-cost text-to-speech synthesis and automated podcast generation using open weights

An open-source Python MCP server built on the Kokoro-82M model. It provides private, local speech generation without API keys or cloud dependencies, featuring audio enhancement, multi-speaker podcast generation, and silence trimming.

Runs local Kokoro-82M speech synthesis models with audio enhancement and script processing.

With Webfuse. Run zero-cost offline speech generation to provide local audio prompts for browser automation testing in Webfuse.

Registered tools

generate_speech(text, voice?, speed?: number, output_file?): Synthesize speech locally using Kokoro voice models.
generate_podcast(segments: array, output_path, gap_duration?: number): Orchestrate multi-speaker podcast dialogues with customizable pacing.
list_voices: List available local voice models and language presets.
process_script(script_path, output_path, voice?): Convert a multi-line text script into sequential audio files.

Install

pip install -e '.[mcp,cli]'

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "kokoro": {
      "command": "aparsoft-tts-server",
      "args": [],
      "env": {
        "TTS_VOICE": "am_michael",
        "TTS_SPEED": "1.0"
      }
    }
  }
}

Limitations

Downloads local model weights on first execution, introducing initial setup latency.
Synthesis speed depends on host CPU or GPU capabilities during batch audio generation.

aparsoft/kokoro-mcp-server

Works withClaude DesktopCursorOpenCode

Standard MCP tools hit Cloudflare and auth walls.

Webfuse Augmented Web Proxy works through bot challenges.

Start Building FreeFree tier · no credit card

AI Agent

MCP

Find the claim form

Fill patient details

Submit claim

Task completed

Local Speech Recognition & Async Jobs

#08 stt-mcp

sm18lr88 31 stdio

Best for: Local, private speech-to-text transcription with asynchronous job lifecycle management and CUDA safety policies

A local speech-to-text MCP server designed for privacy and stability. It features process-isolated workers, fail-closed GPU policies, and asynchronous job tracking to prevent blocking MCP transports during long audio transcriptions.

Provides process-isolated speech-to-text powered by Granite Speech or parakeet.cpp.

With Webfuse. Transcribe customer audio feedback locally and link transcripts to Webfuse session diagnostic reports.

Registered tools

start_transcription(job_id, source_path, formats?: array): Start an asynchronous transcription job returning an immediate status handle.
get_transcription_status(job_id): Check progress, execution state, and output paths of a transcription job.
transcribe(source_path, formats?: array): Transcribe an audio file synchronously and return structured text results.
cancel_transcription(job_id): Cancel an ongoing background transcription task.

Install

uv run stt-mcp

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "stt": {
      "command": "uv",
      "args": ["run", "stt-mcp"]
    }
  }
}

Limitations

Operates one transcription job at a time per worker process, rejecting concurrent requests as busy.
Requires local audio file paths on the host machine rather than remote URLs or base64 streams.

sm18lr88/STT-MCP

Works withClaude DesktopCursorOpenCode

Apple Silicon STT

#09 mlx-whisper-mcp

Kachi Odoemene 24 stdio

Best for: On-device speech-to-text transcription optimized specifically for Apple Silicon hardware

A lightweight Python MCP server using Apple MLX Whisper. It enables fast on-device speech transcription and translation on macOS without cloud API keys, processing local files and base64 audio inputs.

Uses Apple MLX frameworks for fast on-device Whisper Large v3 Turbo transcription.

With Webfuse. Run private, on-device audio transcription on Mac developer machines while testing Webfuse voice agents.

Registered tools

transcribe_file(file_path, language?, task?): Transcribe an audio file into text using on-device MLX Whisper models.
transcribe_audio(audio_data, file_format, language?): Transcribe base64-encoded audio data directly in memory.
transcribe_youtube(url, language?, task?): Fetch and transcribe audio tracks from YouTube URLs.
download_youtube(url, keep_file?: boolean): Download audio streams from video URLs to a local data directory.

Install

uv run mlx_whisper_mcp.py

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "mlx-whisper": {
      "command": "uv",
      "args": ["run", "/path/to/mlx_whisper_mcp.py"]
    }
  }
}

Limitations

Compatible only with Apple Silicon (M-series) Macs running macOS.
Initial execution downloads a 1.6GB Whisper Large v3 model before transcription begins.

kachiO/mlx-whisper-mcp

Works withClaude DesktopCursorOpenCode

Research & Audio Summaries

#10 notebooklm-mcp (PleasePrompto)

PleasePrompto 3.4k stdio

Best for: Document research, citation extraction, and triggering Google NotebookLM deep research audio overviews

The widely adopted community MCP server for Google NotebookLM. It allows coding agents to query source notebooks, manage document attachments, and initiate deep conversational audio overviews.

Connects agents to NotebookLM chat sessions and conversational audio generation tools.

With Webfuse. Gather documentation from protected web pages using Webfuse stealth proxies before submitting to NotebookLM.

Registered tools

ask_question(question, notebook_id?): Ask questions grounded in notebook sources with exact citations.
add_source(notebook_id, url?, text?): Add a web link, text document, or PDF to a research notebook.
generate_audio_overview(notebook_id, instructions?): Trigger a two-speaker conversational audio summary of the notebook sources.
list_notebooks: List all active notebooks and document counts.

Install

npx -y notebooklm-mcp

Configurationclaude_desktop_config.json

{
  "mcpServers": {
    "notebooklm-prompto": {
      "command": "npx",
      "args": ["-y", "notebooklm-mcp"]
    }
  }
}

Limitations

Requires persistent browser cookies and manual Google login setup.
Long audio generation tasks can take several minutes to process on Google servers.

PleasePrompto/notebooklm-mcp npm

Works withClaude DesktopCursorOpenCode

Which server to use

Start from the constraint you have, whether that is transport, host, or network access, and take the server built for it.

01
You need the highest quality speech synthesis, voice cloning, and sound effects with full conversational agent creation
The official vendor-maintained server with the most complete audio synthesis and voice design toolset.
elevenlabs/elevenlabs-mcp
02
You need programmable SMS messaging, telephone call dispatching, and carrier phone number management
Backed by official Twilio OpenAPI specs providing end-to-end access to Twilio communication APIs.
@twilio-alpha/mcp
03
You require ultra-low latency streaming voice responses with accent adaptation and URL-based audio links
Engineered specifically for low-latency streaming TTS, returning clean download URLs rather than context-heavy byte arrays.
cartesia-mcp
04
You want to deploy, test, and dial full conversational voice agents configured with custom LLMs and transcribers
Offers end-to-end assistant configuration, phone number provisioning, and outbound call dispatching.
@vapi-ai/mcp-server
05
You use Retell AI to manage voice bots and need to create web call tokens or place telephone calls
Provides complete control plane coverage for Retell AI assistants, dynamic LLM prompts, and call analytics.
retellai-mcp-server
06
You want to convert research documents and web pages into two-speaker conversational podcast audio
Automates Google NotebookLM workflows to produce multi-speaker audio overviews and citation-backed summaries.
notebooklm-mcp
07
You need zero-cost, completely offline speech synthesis without API keys or cloud dependencies
Runs open-weight Kokoro-82M models locally with multi-speaker script and podcast generation tools.
kokoro-mcp-server
08
You need asynchronous speech-to-text transcription with background job tracking so the MCP transport never hangs
Features process isolation, non-blocking job status handles, and fail-closed GPU management.
stt-mcp
09
You develop on Apple Silicon Macs and want fast, private on-device Whisper transcription
Hardware-accelerated for Apple M-series chips with zero configuration and local MLX execution.
mlx-whisper-mcp
10
You want community-driven Google NotebookLM document grounding and audio overview generation
The most starred community server for automating research queries and audio discussions.
notebooklm-mcp (PleasePrompto)

Every server above automates a browser you cannot see.

Webfuse acts in the live session.

Start Building FreeFree tier · no credit card

AI Agent

MCP

Find the claim form

Fill patient details

Submit claim

Task completed

Frequently asked questions

How do voice AI MCP servers handle large audio payloads without exhausting LLM context tokens?

Servers take different approaches to audio data. Modern streaming servers like cartesia-mcp write audio to persistent storage and return short download URLs, keeping LLM prompts concise. In contrast, servers configured to return raw base64 audio strings in elevenlabs-mcp can quickly consume hundreds of thousands of tokens, making URL-based or file-based output modes strongly preferable for agent workflows.

Can voice AI MCP servers place real telephone calls to customers?

Yes, servers like twilio-mcp, vapi-mcp, retellai-mcp, and elevenlabs-mcp include tools to initiate outbound PSTN telephone calls. Because these tools place live calls and incur carrier or platform charges immediately, teams should implement human-in-the-loop approval checkpoints before granting agents access to outbound dialing tools.

Which voice and speech recognition MCP servers can run completely offline?

Both kokoro-tts-mcp for speech synthesis and stt-mcp or mlx-whisper-mcp for speech-to-text run completely on local hardware without sending data to cloud APIs. This ensures zero API costs and full data privacy for sensitive voice recordings.

How does Webfuse integrate with Voice AI and Telephony MCP servers?

Webfuse provides live web session virtualization and zero-code co-browsing. When combined with voice MCP servers like elevenlabs-mcp or vapi-mcp, AI agents can speak with customers in real time while simultaneously seeing, navigating, and filling out live web interfaces with human approval gates.

Related guides

Developer & Coding MCP Servers

Ten MCP servers for Git, code search, container execution, error tracking, cloud infrastructure, and databases.

Web Scraping & Content Extraction

Ten MCP servers for web scraping, crawling, SERP search, and markdown extraction.

Customer Support & CRM

Ten MCP servers for Zendesk, Jira Service Management, Salesforce, Freshdesk, and CRM workflows.

MCP Complete Reference

The full Model Context Protocol reference: transports, handlers, capability negotiation, and host configuration.

WebMCP Protocol Reference

How WebMCP exposes tools from inside a live browser session, and where it differs from stdio and HTTP MCP.

Agent Skills Standard

The SKILL.md standard for packaging agent workflows, and how skills compose with MCP tool access.

Playwright Automation

Selector strategy, waiting, and tracing for the browser engine most MCP automation servers wrap.