Speak

Speak

Announce plans, issues, and summaries aloud. Triggered automatically after major milestones.

The default path is local and zero-key: call say_tts (macOS, no API key) with voice unset so /usr/bin/say uses the host's configured System Voice. A plain terminal say "text" works this way too, and the MCP tool must preserve that behavior. Cloud voices and explicit installed say voices are optional enhancements when a saved config or the user selects them.

When to Announce

Announce automatically after a milestone — or when you are blocked and need the user:

• Planning complete - a plan or todo list is finalized
• Issue resolved - a bug fix or error is resolved
• Summary generated - a sprint or major task is completed
• Input needed - you have a question, or need a decision, direction, or approval before continuing. The user may be away or running several agents, so speak up so they know to come back.

When NOT to Announce

Stay silent for:

• Trivial or intermediate steps (single edits, mid-task progress, routine tool calls)
• Output that is already short text the user is reading right now
• Non-interactive, headless, or CI runs (no one is listening)
• When the user has asked for quiet, or already heard an announcement for the same milestone
• More than once per logical milestone — do not narrate every step

Identify the Speaker

Many agents may be announcing from different projects and tmux sessions at the same time, so every announcement MUST start by saying who is speaking. Lead with a short spoken label, then the message:

"<speaker> says: <message>"

Determine the label once and cache it as speaker in .claude/tts-config.json:

• Default to the project directory name:

  basename "$(git rev-parse --show-toplevel 2>/dev/null || pwd)"

• Override via the speaker field in config for a custom phrase (e.g. "Claude on the auth service").

Keep the preamble short — it is spoken before every announcement. Good: "mcp-tts says: ...", "On the auth service, ...", "Claude on project X: ...". This identifies the source even when several projects share the same say voice.

Choosing a Provider (fail fast — never probe secrets)

Pick the provider once, cheaply. Do not inspect environment variables or shell startup files to discover credentials, and do not call every cloud provider just to see which one fails.

1. Reuse a saved choice - if .claude/tts-config.json exists, use it and skip detection. 2. No saved choice -> use `say_tts` directly. Leave voice unset unless Voice Identity has intentionally assigned exact installed say voices. This is the common case and must be instant. 3. Use a cloud provider only when the user or saved config selects one. Preference is google, then openai, then elevenlabs, with say_tts as the final fallback. Optionally assign per-message voices (see Voice Identity). 4. On cloud auth/config errors, fall back. Mark that provider unavailable in .claude/tts-config.json and use the next saved provider or say_tts. 5. Persist intentional choices in .claude/tts-config.json so later announcements skip detection.

say_tts with voice unset is the guaranteed last resort and does not require credential discovery.

Workflow

1. Detect the message type — planning, issue, summary, or question (input needed). 2. Pick the provider (above) — cached from .claude/tts-config.json if present, else use say_tts. 3. Transform the text to speech-friendly form (see Text Transformation), and prepend the speaker label (see Identify the Speaker). 4. Call the chosen TTS tool. On error, fall back per the Error Handling table; say_tts is the guaranteed final step.

Error Handling

An announcement is best-effort — never let it block or derail the main task. On any failure, fall back; the final fallback (say_tts with voice unset) needs no key.

Persist auth/config failures. When a provider fails for a missing key, add it to unavailable_providers in .claude/tts-config.json so it is skipped next time:

{
  "provider_order": ["openai", "say"],
  "unavailable_providers": ["google"]
}

Text Transformation

Convert verbose output to conversational speech:

Target length: ~15-30 seconds of speech (roughly 50-100 words).

Tone by type:

• Planning: "Here's the plan..." (forward-looking, organized)
• Issue: "Found a problem..." (alert but calm)
• Summary: "All done..." (satisfied, accomplished)
• Question: "I need your input on..." (direct — clearly state the question or decision, then stop and wait)

A question announcement reuses the issue voice if no dedicated voice is assigned (both signal the user is needed).

TTS Tools

say_tts (default — local, free, no API key)

mcp__mcp-tts__say_tts
- text: string (required)
- voice: string (optional; any installed macOS voice — see `say -v '?'`)
- rate: integer (50-500; recommended 200-250; default 200)

• Prefer leaving voice unset to use the host's configured System Voice. This is required to preserve the same behavior as a plain terminal say "text" command, including Siri System Voices.
• If voice unset is silent while plain say "text" works, treat that as an MCP server/process bug or stale server process, not as a reason to hardcode a downloaded voice.
• If Voice Identity intentionally chooses a say voice, pass only an exact installed name from /usr/bin/say -v '?'.
• Rate hard limit is 50-500; keep 200-250 for comfortable listening, go higher only when the user explicitly asks.

google_tts (cloud, preferred when configured)

mcp__mcp-tts__google_tts
- text: string (required)
- voice: string (default: "Kore")
- model: string (default: "gemini-3.1-flash-tts-preview")

Voices: 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 (cloud fallback)

mcp__mcp-tts__openai_tts
- text: string (required)
- voice: string (default: "alloy") - alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer, verse
- model: string (default: "gpt-4o-mini-tts-2025-12-15") - accepted: gpt-4o-mini-tts-2025-12-15, tts-1, tts-1-hd
- speed: number (0.25-4.0, default: 1.0)
- instructions: string (voice modulation hints)

voice and model are enforced enums — use only the values listed above, do not improvise.

elevenlabs_tts (cloud fallback)

mcp__mcp-tts__elevenlabs_tts
- text: string (required)

• Voice/model are not tool parameters; they come from the server defaults (premade "Sarah") or the ELEVENLABS_VOICE_ID / ELEVENLABS_MODEL_ID env vars.
• Free-tier API keys can only use premade voices. A Voice Library (community/professional) voice returns HTTP 402 paid_plan_required — handle per the Error Handling table.

Voice Identity (optional, cloud or explicit say voices)

Skip this unless distinct per-project or per-message voices are wanted. It exists so each project and message type can be recognized from another room.

For cloud providers: 1. Read references/voice-pools.json for candidate voices per provider and message type. 2. Check ~/.claude/tts-assignments.json for voices already used by other projects (avoid reuse). 3. Pick one voice per message type (planning/issue/summary) from the configured provider's pool. 4. Save to .claude/tts-config.json and record in ~/.claude/tts-assignments.json.

For macOS say: 1. Run /usr/bin/say -v '?' and parse the exact installed voice names. 2. Prefer installed Premium voices first, then Enhanced voices, then stable legacy voices such as Samantha or Alex. 3. Choose different voices for planning, issue/question, and summary only when good matches exist. 4. If a message type has no good installed candidate, leave voice unset for that type so the host System Voice is used. 5. Do not infer that the host default is broken, and do not replace an unset voice as a fallback. Passing a voice is an intentional identity choice only.

Example .claude/tts-config.json:

{
  "speaker": "mcp-tts",
  "provider_order": ["google", "say"],
  "unavailable_providers": [],
  "voices": {
    "planning": { "provider": "google", "voice": "Kore" },
    "issue": { "provider": "google", "voice": "Aoede" },
    "summary": { "provider": "google", "voice": "Charon" }
  }
}

Examples

Each example leads with the speaker label so the listener knows which project/agent is talking.

Planning (after TodoWrite with multiple items):

"mcp-tts says: Here's the plan for the authentication feature. First, I'll create the login component. Then add session management. Finally, write the tests. Three tasks total."

Issue (after fixing an error):

"mcp-tts says: Found and fixed an issue. The rate limiter wasn't catching timeout errors. Added a try-catch block in the handler. Tests are passing now."

Summary (after completing a feature):

"mcp-tts says: All done with the authentication system. Added login, logout, and session management. Created five new files and updated the main router. Ready for review."

Question (blocked, needs a decision):

"mcp-tts says: I need your input. Should the session tokens expire after one hour or stay valid for a day? I'll wait for your call before wiring up the middleware."