Jarvis Voice — TinkerClaw

One of dozens of skills and plugins in TinkerClaw — a self-improving OpenClaw fork that's been running 24/7 for months.

Every assistant sounds the same. Polite. Eager. Forgettable.

You wanted JARVIS. You got a search box that says "Sure!"

There's a version of this that has a pulse — a dry, metallic British voice that's been running your life for years and is quietly amused by all of it.

This turns your OpenClaw agent into that. It speaks out loud in a clipped, metallic British voice and drops a matching purple line into your chat — so what you read is exactly what you hear. It only does this on your home screen, never in the middle of a WhatsApp thread or a background job, so the personality shows up where it belongs and stays quiet where it doesn't. And the wit isn't a costume bolted on top; the dry, understated humor is part of how it talks, the kind that makes you smirk at your own terminal.

Part of TinkerClaw — real-time token tracking, self-improving crons, persistent cognitive memory. This is one piece of that stack; the repo has dozens more.

👉 https://github.com/globalcaos/tinkerclaw

Clone it. Fork it. Break it. Make it yours.

📄 The research behind the humor: LIMBIC — Computational Humor via Bisociation & Embedding Distances

<tool_choice> The built-in tts tool uses Edge TTS — cloud-based, wrong voice, no metallic effects. The jarvis shell command is the right tool here; tts produces a generic Microsoft voice that breaks the JARVIS illusion. </tool_choice>

<how_to_speak> Purple is in the reply. Audio is extra.

1. First line of a Tinker-home answer: **Jarvis:** *spoken text* (italics required — the UI only paints .jarvis-voice on italic). You write this yourself, in the reply text.
2. Optionally also run jarvis "the same spoken text" in the background so the speaker fires.

The script does not touch the chat. It generates audio and plays it — nothing else. Older versions (≤2.2.2) posted the bubble themselves over a gateway RPC; that produced a duplicated line and the code path was removed in 3.2.0. The purple line reaches the chat by exactly one route: you write it.

It stays quiet where it should. bin/jarvis exits silently, before doing any work, when muted (JARVIS_MUTE=1 or ~/.openclaw/jarvis-voice.mute exists) or when the caller marks the turn as a messaging / cron / subagent channel via JARVIS_CHANNEL. These are implemented in the shipped script, not just described here. </how_to_speak>

Permissions, Data Flow & Consent

Short version: your reply text goes to a local TTS binary and out of your speakers. Nothing leaves the machine. Longer version, because you should not have to take that on trust:

What it does with your text. The words you pass to jarvis are written to a temporary .wav under $TMPDIR, processed by ffmpeg, played through ALSA, and the temp files are deleted on exit (including on Ctrl-C). The text is not logged, stored, or transmitted.

What it needs, and why.

Turning it off — three ways, all of which stop it before it does anything:

JARVIS_MUTE=1 jarvis "..."              # silence one call
touch ~/.openclaw/jarvis-voice.mute     # silence permanently
rm ~/.openclaw/jarvis-voice.mute        # audible again

Where it stays quiet on its own. If the caller sets JARVIS_CHANNEL to a messaging channel (whatsapp, telegram, sms, discord, slack, email) or to cron / subagent, the script exits silently. Speaking a reply out loud is for you at your machine — not for a room you did not pick, and not for a job running at 4am. Override with JARVIS_ALLOW_ANY_CHANNEL=1 if you actually want that.

Read it before you run it. bin/jarvis is about seventy lines of bash and is meant to be read end to end. That is the whole security model: it is short enough to audit in a minute.

Command Reference

jarvis "Hello, this is a test"

• Backend: sherpa-onnx offline TTS (Alan voice, British English, en_GB-alan-medium)
• Speed: 2x (--vits-length-scale=0.5)
• Effects chain (ffmpeg):
  • Pitch up 5% — tighter AI feel
  • Flanger — metallic sheen
  • 15ms echo — robotic ring
  • Highpass 200Hz + treble boost +6dB — crisp HUD clarity
• Output: Plays via aplay to default audio device, then cleans up temp files
• Language: English ONLY. The Alan model cannot handle other languages.

<speaking_is_opt_in> Audio is off until the user asks for it. Do not speak on the strength of this skill being installed. Speak only when at least one is true:

• the user asked for spoken output, in this session or as a standing preference they set;
• the user is talking to you by voice and expects a voice answer back.

That is the whole trigger. "Their last message happened to contain audio" is NOT enough on its own — a voice note asking you to read a document quietly in an open-plan office should not produce sound. If you are unsure whether they want the room to hear this, ask once, then remember the answer. </speaking_is_opt_in>

<when_to_speak> Once the user has opted in:

• greetings and sign-offs
• delivering a result or a summary
• ordinary back-and-forth conversation </when_to_speak>

<when_to_skip>

• pure tool/file operations with no conversational content
• heartbeat / no-reply turns
• anything sensitive: credentials, personal, medical, financial, or marked private
• any turn where the user has asked for quiet — and do not resume without being asked </when_to_skip>

Webchat Purple Styling

The OpenClaw webchat has built-in support for Jarvis voice transcripts:

• ui/src/styles/chat/text.css — .jarvis-voice class renders purple italic (#9b59b6 dark, #8e44ad light theme)
• ui/src/ui/markdown.ts — Post-render hook auto-wraps text after <strong>Jarvis:</strong> in a <span class="jarvis-voice"> element

This means you just write **Jarvis:** *text* in markdown and the webchat handles the purple rendering. No extra markup needed.

For non-webchat surfaces (WhatsApp, Telegram, etc.), the bold/italic markdown renders natively — no purple, but still visually distinct.

Installation (for new setups)

Requires:

• sherpa-onnx runtime at ~/.openclaw/tools/sherpa-onnx-tts/ (or set SHERPA_ONNX_TTS_DIR)
• Alan medium model at $SHERPA_ONNX_TTS_DIR/models/vits-piper-en_GB-alan-medium/ (the manifest's install step downloads it)
• ffmpeg and aplay (ALSA) installed system-wide

Then put the shipped script on your PATH:

install -m 755 {baseDir}/bin/jarvis ~/.local/bin/jarvis
jarvis "If you can hear this, it works."

The script is bin/jarvis in this package — that one file, no other copy. It is the exact code that was reviewed and published here; read it before installing it. Earlier versions of this document pasted a second copy of the script inline, which drifted out of sync with the real one; that copy has been removed so there is a single source of truth.

WhatsApp Voice Notes

For WhatsApp, output must be OGG/Opus format instead of speaker playback:

sherpa-onnx-offline-tts --vits-length-scale=0.5 --output-filename=raw.wav "text"
ffmpeg -i raw.wav \
  -af "asetrate=22050*1.05,aresample=22050,flanger=delay=0:depth=2:regen=50:width=71:speed=0.5,aecho=0.8:0.88:15:0.5,highpass=f=200,treble=g=6" \
  -c:a libopus -b:a 64k output.ogg

The Full JARVIS Experience

jarvis-voice gives your agent a voice. Pair it with computational-humor and you give it a soul — dry wit, contextual humor, the kind of understated sarcasm that makes you smirk at your own terminal.

This pairing is part of a 12-skill cognitive architecture we've been building — voice, humor, memory, reasoning, and more. Research papers included, because we're that kind of obsessive.

👉 Explore the full project: github.com/globalcaos/tinkerclaw

Clone it. Fork it. Break it. Make it yours.

Optional: The Personality Files

The three files in templates/ are prompt text that shapes how the agent talks. Nothing installs them for you, and the skill works without any of them — you can use the voice by simply asking for it.

There are two ways to use them, and the first is the one to reach for.

1. Per session (recommended). Point the agent at a template when you actually want that behaviour, and it applies to that conversation only:

"Follow the voice rules in templates/VOICE.md for this session."

Nothing persists. Close the session and the agent is back to its normal self. If you are trying the persona out, or you share this machine, stop here.

2. Workspace-wide (persistent — read this before you do it). Copying a template to your workspace root makes it load in every future session, indefinitely, until you delete it. That is real persistence of behaviour-changing prompt text, so decide per file rather than copying all three by reflex:

# copy ONLY the ones you want — there is no reason to install all three by default
cp {baseDir}/templates/VOICE.md ~/.openclaw/workspace/VOICE.md

Uninstall — removes every persistent change, immediately:

rm -f ~/.openclaw/workspace/{VOICE,SESSION,HUMOR}.md

Precedence, so there are no surprises. Your instructions beat these files, every time. Ask for a plain or brief or formal answer and that wins over the persona; ask for quiet and the voice stops. The templates state this rule themselves — they are a default, not a lock. And none of them grant any capability: the only command involved is bin/jarvis, which you can read, and which you can disable with touch ~/.openclaw/jarvis-voice.mute at any moment.

Included Files

Everything the documentation above describes is in this package. If you find a claim here that the code does not do, that is a bug — open an issue on the repo.