- Home
- Skills
- Agents & Orchestration
- Speak: Offline Voice Alerts for AI Agents
Works with the AI Agents you already use
Speak: Offline Voice Alerts for AI Agents
Your agent says build results, test failures, and questions out loud, using a text-to-speech engine that runs on your own CPU.
$6
Speak: Offline Voice Alerts for AI Agents
Example session with this skill installed
Set up speak so you tell me out loud when a long task finishes or fails. Use the eponine voice, keep it short, then run one test and tell me how it went.
- Read your context and instructions
- Compiled the speak
Summary
Speech is on. I will speak when a task completes or fails, in short conversational lines, using the eponine voice. One test line rendered in about 6 seconds on a low-end laptop.
Settings saved to ./.speak.md
---
engine: pocket-tts
server_url: http://localhost:8000
voice: eponine
tone: conversational
frequency: completions-and-errors
tasks:
builds_and_tests: true
task_completion: true
errors_and_crashes: true
research_findings: true
code_reviews: false
audio:
volume: 1.5
mode: safe
---
Test run
powershell -ExecutionPolicy Bypass -File "scripts/speak.ps1" `
-Text "Build complete. All tests passed, ready to ship." `
-Voice eponine -OutFile test.wav -NoPlay
| Check | Result |
|---|---|
| Exit code | 0 |
| First run | 24.8 s (uv built its Python environment once) |
| Later runs | about 6 s |
| Audio | 24 kHz, mono, about 4.0 s |
| Server | Pocket-TTS on localhost:8000, status healthy |
I used -NoPlay -OutFile so this test wrote a file instead of playing sound. Without those flags, the line plays through your speakers.
Next steps
- Say "only speak when builds fail" to narrow it.
- Say "switch your voice to michael" to change the voice.
- Say "mute speak" to silence it, and "unmute speak" to bring it back.
Connects securely to your tools. The creator never sees your data.
About this skill
Hear your agent finish, fail, or ask. No terminal watching.
Speak gives your AI coding agent a voice. A small text-to-speech model runs on your own CPU. It needs no API key and no account. Your text and audio stay on your machine.
Start a long build, switch windows, and keep working. When the build finishes or breaks, your agent tells you.
What you get
- Rules that tell your agent when to speak and when to stay quiet.
- Launcher scripts for Windows (PowerShell) and macOS or Linux (Bash). Only Windows is verified.
- A setup guide that installs the speech engine and plays voice samples.
- A settings guide: change voice, tone, frequency, volume, or mute.
- Three ready-made setups, example wording for each tone, and a troubleshooting table.
- Direct commands for agents that only receive the skill text.
What the agent says
| Moment | Example |
|---|---|
| Task start | "Starting the database migration." |
| Milestone | "Schema updated. Running tests." |
| Completion | "Build complete. All tests passed, ready to ship." |
| Error | "Tests failed in the auth module." |
| Key finding | "Found the cause. A missing null check in the parser." |
Messages are one or two sentences. The agent never reads code, diffs, or stack traces aloud.
What you control
Ask in plain words, for example "only speak when builds fail".
| Setting | Options |
|---|---|
| Tone | conversational, formal, minimal |
| Frequency | completions-and-errors, all-milestones, errors-only, silent |
| Voice | Any voice from your engine, such as eponine, michael, javert |
| Volume | A multiplier. 1.0 is default. 2.5 is for noisy rooms. |
| Task types | Builds and tests, reviews, research findings, completions, errors |
How it works
- The agent reads
.speak.mdin your repo, or~/.config/speak/config.md. - It runs a launcher script in the background. Your agent never waits for audio.
- The launcher checks that the speech server is up. If it is down, it starts Pocket-TTS.
- A lock file lets only one agent speak at a time. This matters when several agents finish together.
- If anything fails, the script exits quietly. Speech never breaks your real task.
Engines
| Engine | Status | Who starts it |
|---|---|---|
| Pocket-TTS (Kyutai Labs) | Default. Conversational voice. Verified on Windows 11. | The skill starts it for you. |
| Kokoro-82M (hexgrad) via Docker | Supported by the scripts. Not yet verified. | You start the container once. |
What you need
uv(installs Python for you if it is missing)- About 1.5 GB of free RAM for Pocket-TTS
- Speakers or headphones
- Internet for the first run only: Python packages and the model download once
Measured on a low-end laptop
Intel Core i3-10110U, 8 GB RAM, Windows 11, on AC power.
| Measure | Result |
|---|---|
| Server start from a fresh restart, model on disk | 31 s |
| One 17-word request to the server | 3.6 to 4.8 s |
| One 68-word request | 22.2 s |
| Whole launcher call, 8 words | about 6 s (24.8 s on the very first run) |
| Memory the server uses | 1.3 GB committed, 0.8 GB resident |
| Audio produced | 24 kHz mono |
Speech arrives a few seconds after the event. It is not instant on a slow CPU.
Limits
- Verified on Windows 11. The macOS and Linux scripts are included but not yet verified.
- Run the setup step once. It downloads the model, which takes longer than a normal call.
- Speech needs a local server running. The skill starts Pocket-TTS if it is installed.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Requires local TTS server installation. Audio playback fails silently if no audio device is detected. Messages are limited to short summaries to avoid playback lag.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 26 days ago
- Software
uv (Python tool runner)
Speakersorheadphonespocket-ttsfound- Optional
Docker
The skill handles installation; pocket-tts installs through uv. Verified on Windows. Docker is only for the Kokoro engine.
Permissions
File Scopes
Terminal: starts the local speech server and audio player. Network: localhost only while speaking; the first run downloads packages and model weights once. Windows: -ExecutionPolicy Bypass runs the bundled local script. stream_tts.py raises its own process priority on Windows (ctypes SetPriorityClass) so audio does not crackle, and starts the speech server with subprocess.Popen, passing a copy of the environment plus HF_HUB_OFFLINE=1. The scripts read no credentials. No text or audio leaves the machine.