mcp-tts

<p align="center">
  <a href="https://github.com/blacktop/mcp-tts"><img alt="mcp-tts Logo" src="https://raw.githubusercontent.com/blacktop/mcp-tts/main/docs/logo.webp" height="200" /></a>
  <h1 align="center">mcp-tts</h1>
  <h4><p align="center">MCP Server for TTS (Text-to-Speech)</p></h4>
  <p align="center">
    <a href="https://mcptoplist.com/server/mcp.so%2Fmcp-tts%2Fblackto" alt="Actions">
          <img src="https://mcptoplist.com/badge/mcp.so%2Fmcp-tts%2Fblacktop.svg" /></a>
    <a href="https://github.com/blacktop/mcp-tts/actions" alt="Actions">
          <img src="https://github.com/blacktop/mcp-tts/actions/workflows/go.yml/badge.svg" /></a>
    <a href="https://github.com/blacktop/mcp-tts/releases/latest" alt="Downloads">
          <img src="https://img.shields.io/github/downloads/blacktop/mcp-tts/total.svg" /></a>
    <a href="https://github.com/blacktop/mcp-tts/releases" alt="GitHub Release">
          <img src="https://img.shields.io/github/release/blacktop/mcp-tts.svg" /></a>
    <a href="http://doge.mit-license.org" alt="LICENSE">
          <img src="https://img.shields.io/:license-mit-blue.svg" /></a>
</p>
<br>

## What? 🤔

Adds Text-to-Speech to things like Claude Desktop and Cursor IDE.  

It registers the existing TTS tools, plus `voice_tts` when `voice-say` is available on `PATH`:
 - `say_tts` 
 - `voice_tts` (when available)
 - `elevenlabs_tts`
 - `google_tts`
 - `openai_tts`

### `say_tts`

Uses the macOS `say` binary to speak the text with built-in system voices

### `voice_tts`

Uses the local `voice-say` CLI from [Voice](https://github.com/blacktop/Voice) to run Qwen3-TTS through MLX. Voice needs no API key and keeps text and audio on the Mac. The model reloads for every invocation, so startup typically takes several seconds; prefer it for summaries and announcements rather than time-critical alerts.

The tool is registered only when `voice-say` is resolvable on `PATH`. Optional parameters are:

- `voice`: preset voice `Ryan` or `Aiden`
- `tier`: `small` for faster 0.6B synthesis or `large` for higher-quality 1.7B synthesis
- `style`: free-text delivery guidance
- `describe`: free-text voice design; forces the 1.7B model and cannot be combined with `voice`

Voice plays audio directly and does not support `--output-dir`; with `--no-play`, calls fail without launching `voice-say`.

> [!WARNING]
> The Voice repository is currently private and will be released publicly soon, so the link requires access for now.

### `elevenlabs_tts`

Uses the [ElevenLabs](https://elevenlabs.io/app/speech-synthesis/text-to-speech) text-to-speech API to speak the text with premium AI voices

### `google_tts`

Uses Google's [Gemini TTS models](https://ai.google.dev/gemini-api/docs/speech-generation) to speak the text with 30 high-quality voices. Available voices include:

**Achernar, Achird, Algenib, Algieba, Alnilam, Aoede, Autonoe, Callirrhoe, Charon, Despina, Enceladus, Erinome, Fenrir, Gacrux, Iapetus, Kore, Laomedeia, Leda, Orus, Puck, Pulcherrima, Rasalgethi, Sadachbia, Sadaltager, Schedar, Sulafat, Umbriel, Vindemiatrix, Zephyr, Zubenelgenubi**

### `openai_tts`

Uses OpenAI's [Text-to-Speech API](https://platform.openai.com/docs/guides/text-to-speech) to speak the text with 10 natural-sounding voices:

- **alloy** (Warm, conversational, modern)
- **ash** (Confident, assertive, slightly textured)
- **ballad** (Gentle, melodious, slightly lyrical)
- **coral** (Cheerful, fresh, upbeat)
- **echo** (Neutral, calm, balanced)
- **fable** (Storyteller-like, expressive)
- **nova** (Clear, precise, slightly formal)
- **onyx** (Deep, authoritative, resonant)
- **sage** (Soothing, empathetic, reassuring)
- **shimmer** (Bright, animated, playful)
- **verse** (Versatile, expressive)

Supports three quality models:
- **gpt-4o-mini-tts** - Default, optimized quality and speed
- **tts-1** - Standard quality, faster generation  
- **tts-1-hd** - High definition audio, premium quality

Additional features:
- Speed control from 0.25x to 4.0x (default: 1.0x)
- Custom voice instructions (e.g., "Speak in a cheerful and positive tone") via parameter or `OPENAI_TTS_INSTRUCTIONS` environment variable

## Configuration

### Sequential vs Concurrent TTS

By default, the TTS server enforces sequential speech operations - only one TTS request can play audio at a time. This prevents multiple agents from speaking simultaneously and creating an unintelligible cacophony. Subsequent requests will wait in a queue until the current speech completes.

**Multi-Instance Protection**: The mutex works both within a single MCP server process and across multiple Claude Desktop instances. When running multiple Claude Desktop terminals, they coordinate via a system-wide file lock to prevent overlapping speech.

To allow concurrent TTS operations (multiple speeches playing simultaneously):

**Environment Variable:**
```bash
export MCP_TTS_ALLOW_CONCURRENT=true
```

**Command Line Flag:**
```bash
mcp-tts --sequential-tts=false
```

> **Note:** Concurrent TTS may result in overlapping audio that's difficult to understand. Use this option only when you explicitly want multiple TTS operations to run simultaneously.

### Suppressing "Speaking:" Output

By default, TTS tools return a message like "Speaking: [text]" when speech completes. This can interfere with LLM responses. To suppress this output:

**Environment Variable:**
```bash
export MCP_TTS_SUPPRESS_SPEAKING_OUTPUT=true
```

**Command Line Flag:**
```bash
mcp-tts --suppress-speaking-output
```

When enabled, tools return "Speech completed" instead of echoing the spoken text.

### Saving Audio to Disk

Save TTS audio output to files instead of (or in addition to) playing them:

**Environment Variables:**
```bash
export MCP_TTS_OUTPUT_DIR=/path/to/audio    # Save audio files to this directory
export MCP_TTS_NO_PLAY=true                  # Skip playback, only save (optional)
```

**Command Line Flags:**
```bash
mcp-tts --output-dir /path/to/audio          # Save and play
mcp-tts --output-dir /path/to/audio --no-play  # Save only, no playback
```

Files are saved with unique names: `tts_{timestamp}_{hash}.{ext}`

| Provider | Format |
|----------|--------|
| macOS say | AIFF |
| Voice | Playback only |
| ElevenLabs | MP3 |
| Google TTS | WAV |
| OpenAI TTS | MP3 |

## Getting Started

### Install

```bash
go install github.com/blacktop/mcp-tts@latest
```

```bash
❱ mcp-tts --help

TTS (text-to-speech) MCP Server.

Provides multiple text-to-speech services via MCP protocol:

• say_tts - Uses macOS built-in 'say' command (macOS only)
• voice_tts - Uses local Qwen3-TTS through voice-say (when available on PATH)
• elevenlabs_tts - Uses ElevenLabs API for high-quality speech synthesis
• google_tts - Uses Google's Gemini TTS models for natural speech
• openai_tts - Uses OpenAI's TTS API with various voice options

Each tool supports different voices, rates, and configuration options.
Requires appropriate API keys for cloud-based services.

Designed to be used with the MCP (Model Context Protocol).

Usage:
  mcp-tts [flags]

Flags:
  -h, --help                       help for mcp-tts
      --no-play                    Skip playback, only save (requires --output-dir)
      --output-dir string          Save audio files to directory (env: MCP_TTS_OUTPUT_DIR)
      --sequential-tts             Enforce sequential TTS (prevent concurrent speech) (default true)
      --suppress-speaking-output   Suppress 'Speaking:' text output
  -v, --verbose                    Enable verbose debug logging
```

### Configuration

#### [Claude Desktop](https://claude.ai/download)

Add to `~/Library/Application Support/Claude/claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "say": {
      "command": "mcp-tts",
      "env": {
        "ELEVENLABS_API_KEY": "********",
        "ELEVENLABS_VOICE_ID": "EXAVITQu4vr4xnSDxMaL",
        "GOOGLE_AI_API_KEY": "********",
        "OPENAI_API_KEY": "********",
        "OPENAI_TTS_INSTRUCTIONS": "Speak in a cheerful and positive tone",
        "MCP_TTS_SUPPRESS_SPEAKING_OUTPUT": "true",
        "MCP_TTS_ALLOW_CONCURRENT": "false"
      }
    }
  }
}
```

#### [Claude Code](https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/overview)

```bash
claude mcp add say \
  -e GOOGLE_AI_API_KEY=your_key \
  -e ELEVENLABS_API_KEY=your_key \
  -e OPENAI_API_KEY=your_key \
  -- mcp-tts
```

#### [Codex CLI](https://platform.openai.com/docs/guides/mcp)

```bash
codex mcp add say \
  --env GOOGLE_AI_API_KEY=your_key \
  --env ELEVENLABS_API_KEY=your_key \
  --env OPENAI_API_KEY=your_key \
  -- mcp-tts
```

#### [Gemini CLI](https://github.com/google/gemini-cli)

```bash
gemini mcp add say mcp-tts \
  -e GOOGLE_AI_API_KEY=your_key \
  -e ELEVENLABS_API_KEY=your_key \
  -e OPENAI_API_KEY=your_key
```

Or manually add to `~/.gemini/settings.json` (or `.gemini/settings.json` in project root):

```json
{
  "mcpServers": {
    "say": {
      "command": ["mcp-tts"],
      "env": {
        "GOOGLE_AI_API_KEY": "..."
      }
    }
  }
}
```

#### Environment Variables

- `ELEVENLABS_API_KEY`: Your ElevenLabs API key (required for `elevenlabs_tts`)
- `ELEVENLABS_VOICE_ID`: ElevenLabs voice ID (optional, defaults to the premade "Sarah" voice `EXAVITQu4vr4xnSDxMaL`). Free-tier API keys can only use **premade** voices — Voice Library (community/professional) voices return `402 paid_plan_required`.
- `GOOGLE_AI_API_KEY` or `GEMINI_API_KEY`: Your Google AI API key (required for `google_tts`)
- `OPENAI_API_KEY`: Your OpenAI API key (required for `openai_tts`)
- `OPENAI_TTS_INSTRUCTIONS`: Custom voice instructions for OpenAI TTS (optional, e.g., "Speak in a cheerful and positive tone")
- `MCP_TTS_SUPPRESS_SPEAKING_OUTPUT`: Set to "true" to suppress "Speaking:" output (optional)
- `MCP_TTS_ALLOW_CONCURRENT`: Set to "true" to allow concurrent TTS operations (optional, defaults to sequential)
- `MCP_TTS_OUTPUT_DIR`: Directory to save audio files (optional)
- `MCP_TTS_NO_PLAY`: Set to "true" to skip playback when saving (optional, requires `MCP_TTS_OUTPUT_DIR`)
- `MCP_TTS_ELICIT`: Set to "true" to enable interactive elicitation prompts (also `--elicit`; optional, default off). When off, tools use explicit arguments/defaults without prompting — recommended for agent use.

### Test

#### Test macOS TTS
```bash
❱ cat test/say.json | go run main.go --verbose

2025/03/23 22:41:49 INFO Starting MCP server name="Say TTS Service" version=1.0.0
2025/03/23 22:41:49 DEBU Say tool called request="{Request:{Method:tools/call Params:{Meta:<nil>}} Params:{Name:say_tts Arguments:map[text:Hello, world!] Meta:<nil>}}"
2025/03/23 22:41:49 DEBU Executing say command args="[--rate 200 Hello, world!]"
2025/03/23 22:41:49 INFO Speaking text text="Hello, world!"
```
```json
{"jsonrpc":"2.0","id":3,"result":{"content":[{"type":"text","text":"Speaking: Hello, world!"}]}}
```

#### Test Google TTS
```bash
❱ cat test/google_tts.json | go run main.go --verbose

2025/05/23 18:26:45 INFO Starting MCP server name="Say TTS Service" version=""
2025/05/23 18:26:45 DEBU Google TTS tool called request="{...}"
2025/05/23 18:26:45 DEBU Generating TTS audio model=gemini-3.1-flash-tts-preview voice=Kore text="Hello! This is a test of Google's TTS API. How does it sound?"
2025/05/23 18:26:49 INFO Playing TTS audio via beep speaker bytes=181006
2025/05/23 18:26:53 INFO Speaking via Google TTS text="Hello! This is a test of Google's TTS API. How does it sound?" voice=Kore
```
```json
{"jsonrpc":"2.0","id":4,"result":{"content":[{"type":"text","text":"Speaking: Hello! This is a test of Google's TTS API. How does it sound? (via Google TTS with voice Kore)"}]}}
```

#### Test OpenAI TTS
```bash
❱ cat test/openai_tts.json | go run main.go --verbose

2025/05/23 19:15:32 INFO Starting MCP server name="Say TTS Service" version=""
2025/05/23 19:15:32 DEBU OpenAI TTS tool called request="{...}"
2025/05/23 19:15:32 DEBU Generating OpenAI TTS audio model=tts-1 voice=nova speed=1.2 text="Hello! This is a test of OpenAI's text-to-speech API. I'm using the nova voice at 1.2x speed."
2025/05/23 19:15:34 DEBU Decoding MP3 stream from OpenAI
2025/05/23 19:15:34 DEBU Initializing speaker for OpenAI TTS sampleRate=22050
2025/05/23 19:15:36 INFO Speaking text via OpenAI TTS text="Hello! This is a test of OpenAI's text-to-speech API. I'm using the nova voice at 1.2x speed." voice=nova model=tts-1 speed=1.2
```
```json
{"jsonrpc":"2.0","id":5,"result":{"content":[{"type":"text","text":"Speaking: Hello! This is a test of OpenAI's text-to-speech API. I'm using the nova voice at 1.2x speed. (via OpenAI TTS with voice nova)"}]}}
```

#### Test the mutex behavior with multiple TTS requests

```bash
# Sequential mode (default) - speeches play one after another
cat test/sequential.json | go run main.go --verbose

# Concurrent mode - allows overlapping speech  
cat test/sequential.json | go run main.go --verbose --sequential-tts=false
```

## Skill: `speak`

This repo includes a **speak** skill that automatically announces plans, issues, and summaries aloud using TTS. It defaults to the host's macOS `say` voice, and can optionally assign distinct voices per project or message type so you can identify what is speaking from another room.

Skills follow the [Agent Skills](https://agentskills.io) open standard and work across Claude Code, Codex CLI, and Gemini CLI.

### Install Skill

#### skills.sh

```bash
npx skills add https://github.com/blacktop/mcp-tts --skill speak
```

#### Claude Code

**Via Plugin Marketplace** (recommended):
```bash
claude plugin marketplace add blacktop/mcp-tts
claude plugin install speak@mcp-tts
```

**Or manually:**
```bash
mkdir -p ~/.claude/skills
git clone https://github.com/blacktop/mcp-tts.git /tmp/mcp-tts
cp -r /tmp/mcp-tts/skill ~/.claude/skills/speak
```

The skill is now available. Claude will use it automatically when relevant, or invoke directly with `/speak`.

#### Codex CLI

**Using the skill-installer** (within a Codex session):
```
$skill-installer install the speak skill from https://github.com/blacktop/mcp-tts --path skill
```

**Or manually:**
```bash
mkdir -p ~/.codex/skills
git clone https://github.com/blacktop/mcp-tts.git /tmp/mcp-tts
cp -r /tmp/mcp-tts/skill ~/.codex/skills/speak
```

Restart Codex after installing.

#### Gemini CLI

Gemini CLI uses **extensions** to bundle skills. Install this repo as an extension:

```bash
gemini extensions install https://github.com/blacktop/mcp-tts.git
```

This installs the `speak` skill.

**Or manually** (skill only):
```bash
mkdir -p ~/.gemini/skills
git clone https://github.com/blacktop/mcp-tts.git /tmp/mcp-tts
cp -r /tmp/mcp-tts/skill ~/.gemini/skills/speak
```

> **Note:** Gemini CLI skills are experimental. Enable via `/settings` → search "Skills" → toggle on.

### Shared Skills Directory (Optional)

To maintain one copy across all agents, run the install script:

```bash
git clone https://github.com/blacktop/mcp-tts.git
cd mcp-tts
./install-skill.sh
```

This copies the skill to `~/.agents/skills/speak` and creates symlinks for Claude Code, Codex CLI, and Gemini CLI.

### Verify Installation

| Agent | Command |
|-------|---------|
| Claude Code | Ask "What skills are available?" or type `/speak` |
| Codex CLI | Skills load automatically on restart |
| Gemini CLI | `gemini extensions list` or check `/settings` for skills |

### How It Works

The skill should be selected after:

- **Planning complete** - When a plan/todo list is finalized
- **Issue resolved** - When a bug fix or error is resolved
- **Summary generated** - When completing a major task
- **Long operation changes state** - When a build, test, deployment, release, research, or monitoring phase completes or fails
- **Human intervention needed** - When approval, authentication, manual action, or an exhausted retry budget blocks progress

The skill defaults to local-only speech without probing credentials: `voice_tts` for plans and summaries when Voice is registered, and fast `say_tts` for urgent alerts or fallback. Both avoid cloud API quotas and keep spoken content on the Mac. Cloud TTS is used only after an explicit user choice or saved configuration. If that cloud provider hits a quota, token, authentication, or configuration failure, automatic cloud TTS is disabled for the session and the skill uses local speech—or remains text-only when no local tool is available. It never tries the other cloud providers automatically. One-off provider corrections last for the session; configuration is updated only when the user asks to remember the choice. For local `say` identity, leave `voice` unset to use the host System Voice unless the user intentionally selected an exact installed voice.

## License

MIT