SecretVoice

Direction lives in the script, not in a settings object.

Every other voice MCP tool takes text and returns audio. That is a utility. This one lets the agent act: it can whisper on one line, sigh on the next, switch to a different character, and answer in another language — without a second tool call.

Emoji	Direction	Emoji	Direction
👂	whisper	💋	kiss
😮💨	sigh	🤭	chuckle
🌬️	breathless	🥺	trembling
🥵	panting	😭	sobbing
⏸️	pause	🐢	slowly

The agent writes the performance. A prompt like “read this line like you are telling a secret at 2am” becomes 👂 read this line… — and the voice actually does it. Any emoji that is not a cue is stripped, so scripts stay readable.

Six tools, three resources.

Everything an agent needs to pick a voice, perform, and stay inside its credit budget.

Tool	What it does	Scope
list_voices	Every voice with id and gender	voices:read
list_languages	The 22 supported language codes	voices:read
text_to_speech	Text in, audio out — emoji cues allowed	tts:generate
generate_scene	Multi-voice scene: one clip per segment, own voice and language	tts:generate
get_balance	Plan and remaining credits	account:read
list_history	Recent generations	history:read

Resources for clients that read context: secretvoice://voices, secretvoice://languages, secretvoice://vocal-cues.

Two lines in a config file.

The client registers itself, you approve the scopes once, and it receives a token bound to this server. Nothing to paste, nothing to rotate.

{
  "mcpServers": {
    "secretvoice": {
      "type": "http",
      "url": "https://api.secretvoice.dev/mcp"
    }
  }
}

{
  "mcpServers": {
    "secretvoice": { "url": "https://api.secretvoice.dev/mcp" }
  }
}

{
  "servers": {
    "secretvoice": { "type": "http", "url": "https://api.secretvoice.dev/mcp" }
  }
}

Agent instructions that actually work. Tell the model what it may do with the voice, not just that a tool exists. This is the block we use:

You have SecretVoice over MCP. When the user wants audio, narration or a scene:
1. Call list_voices and pick by gender and style — never guess a voice id.
2. Write the script with emoji vocal cues where the delivery should change
   (👂 whisper, 😮‍💨 sigh, 🌬️ breathless, 💋 kiss, ⏸️ pause).
3. Use text_to_speech for one speaker, generate_scene for dialogue.
4. State the character count before generating: 1 credit = 1 character.

Client without dynamic registration? Mint a key in the dashboard and send it as Authorization: Bearer <key> — the endpoint accepts both.

Same credits as the API. Nothing hidden.

One credit is one character. A 500-character paragraph is 500 credits, whether it comes from a script, the Studio, or an agent. Plans start at 180,000 credits every month.

Before you connect.

• Is this an official Anthropic server? No — it is an independent MCP server published by SecretVoice.
• Do I need an API key? No. The client registers itself and you approve the scopes once. A key still works for clients that cannot do OAuth.
• Which languages? Twenty-two, including Turkish, Japanese, Korean, Hindi, Tamil and Thai. Ask the server with list_languages.
• Can it do several characters in one scene? Yes — generate_scene takes one segment per line, each with its own voice and language.
• Does it stream audio? Tool calls return complete clips. For sub-200ms live conversation use the WebSocket API.
• How do I revoke it? From your account, or via the revocation endpoint.