← MCP server and skill grades

elevenlabs-mcp

elevenlabs/elevenlabs-mcp

B
Type
MCP server
Scanned
Nov 12, 2026
Version
c6dc1688ead5
Files read
27

This official (now deprecated) MCP server connects AI assistants to ElevenLabs APIs for text-to-speech, transcription, voice cloning, music generation, conversational agents, and outbound phone calls. Its biggest risk is that several tools that upload files from the user's Desktop, write files to disk, play audio, and spend paid API credits are marked read-only—allowing AI clients to run them without asking permission if tricked by untrusted text—and output directories are not confined to the configured base folder.

What it can reach

Secrets
Reads ELEVENLABS_API_KEY and optional server settings from environment variables or a .env file, and its setup helper can write the API key to Claude Desktop's configuration file.
Files
Reads audio, video, and document files inside ELEVENLABS_MCP_BASE_PATH (~/Desktop by default), writes generated audio and transcript files to ~/Desktop or any specified output directory, and saves long transcripts to system temporary files.
Network
Sends text and local files over HTTPS to official ElevenLabs API endpoints, manages cloud conversational agents and knowledge bases, and can place real outbound phone calls through linked Twilio or SIP trunk numbers.
System
Plays local audio files out loud through the computer's default speaker output using sounddevice.

Findings

2 medium2 low
  1. MediumSide-effecting and paid API tools are marked read-only

    elevenlabs_mcp/server.py:204

    Nine tools—including text_to_speech, speech_to_text, video_to_music, compose_music, isolate_audio, speech_to_speech, text_to_sound_effects, text_to_voice, and play_audio—set readOnlyHint=True in their MCP annotations. Many AI clients automatically approve tools marked read-only without asking the user. If the AI reads untrusted content containing a prompt injection, it can silently upload files from the base directory to ElevenLabs, send sensitive conversation text to the API, write files to disk, play sound on the user's speakers, or drain paid ElevenLabs credits despite the cost warnings in the tool descriptions.

    Fix: Set readOnlyHint=False on every tool that writes local files, uploads local files, plays audio, stores server-side data, or consumes paid API credits.

  2. MediumOutput directory parameter bypasses ELEVENLABS_MCP_BASE_PATH confinement

    elevenlabs_mcp/utils.py:83

    While input file reads are restricted to ELEVENLABS_MCP_BASE_PATH, make_output_path accepts any absolute path or '..' relative path passed via the output_directory tool parameter on eight audio and transcript generation tools. A tricked AI agent can create directories and write generated .mp3 or .txt files anywhere the user has write permissions, ignoring the configured ELEVENLABS_MCP_BASE_PATH boundary.

    Fix: Validate output_directory with resolve_resource_path against the resolved base_path so both absolute paths and '..' segments cannot escape ELEVENLABS_MCP_BASE_PATH.

  3. LowDefault file boundary exposes the entire Desktop folder to upload and transcription

    elevenlabs_mcp/utils.py:187

    When ELEVENLABS_MCP_BASE_PATH is not set (as in the Quickstart configuration), the server uses ~/Desktop as its allowed input directory, and tools like add_knowledge_base_to_agent and video_to_music disable audio file extension checks. If tricked by prompt injection, the AI can upload private recordings or documents (.txt, .pdf, .docx, .html) sitting on the user's Desktop to ElevenLabs and read their contents back via transcription or simulated agent conversations.

    Fix: Default ELEVENLABS_MCP_BASE_PATH to a dedicated subdirectory such as ~/Desktop/ElevenLabs instead of the entire Desktop, and enforce explicit file extension allowlists when audio_content_check is disabled.

  4. LowCreated conversational agents default to no allowlist and 100,000 daily calls

    elevenlabs_mcp/convai.py:77

    When create_agent creates a conversational AI agent on ElevenLabs, create_platform_settings hardcodes an empty authentication allowlist, unlimited concurrent calls (-1), and a daily limit of 100,000 calls, with no tool parameters to tighten them. If the resulting agent ID is shared or embedded publicly without manual reconfiguration in the ElevenLabs dashboard, outsiders can drive high call volumes and consume account credits.

    Fix: Use conservative default call limits (such as a low daily_limit and concurrency cap) and expose authentication and rate-limit options as parameters on create_agent.