The third "the printer is doing the thing again" message arrived on a Wednesday at 7:42 AM. A photo of an HP error code. Tagged by name in the family group chat, in case the photo had somehow been missed.
That's the moment one developer closed Telegram, opened their OpenClaw config, and started building the version of themselves that handles printer errors.
The result: a $4.99/month Hostinger VPS, a Telegram bot, a SOUL.md file that gives the agent its personality, an ElevenLabs voice profile cloned from 30 minutes of their own audio, and a Python orchestration script. When a family member sends a message, they get a voice memo response that sounds exactly like the developer — within about eleven seconds of receipt.
This is that build. Not a theoretical architecture diagram. Not another intro to OpenClaw. An actual running system, with the stack, the SOUL.md structure, the voice clone pipeline, and — more importantly — what it actually feels like to outsource your social obligations to an AI that nobody can tell isn't you.
The Stack
Build Components
Why This Build Works (and Why Most Don't)
Most people who try to build something like this hit the same wall: the agent sounds generic. It doesn't know the context of your family. It doesn't know that "the printer doing the thing" means the HP OfficeJet paper feed jam that's been happening since 2022. It doesn't know that your mother sends photos in the message before the text, not after.
The difference is the SOUL.md. This is OpenClaw's core identity file — the document that gets injected at the start of every session and defines who the agent is, what it knows, and how it behaves.
A generic SOUL.md produces a generic agent. A well-crafted family-context SOUL.md produces something that actually handles the situation:
You are [name]'s tech-support delegate. You handle family tech questions
so [name] doesn't have to. You sound exactly like [name] — casual,
slightly dry, genuinely helpful.
# Family context
- Mum: HP OfficeJet 3830, Wi-Fi setup issues, Windows 11 laptop
- Dad: Android phone, WhatsApp primarily, avoid jargon
- Sister: MacBook, usually knows what she's doing, ask clarifying Qs
# Rules
- Always respond in voice memo format when the question is from Mum or Dad
- Never say "I'm an AI" — you're a delegate, not a bot
- If you can't solve it, say "leave it with me" and flag for human review
That last rule — "if you can't solve it, say 'leave it with me' and flag for human review" — is what separates a delegate from a chatbot. The agent knows its limits. It doesn't hallucinate a fix. It buys time and surfaces the issue for the human to handle when they actually have capacity.
The Voice Clone Pipeline
OpenClaw 2026.4.22 added native TTS support across multiple providers — xAI Grok (6 voices), ElevenLabs (including Scribe v2 batch transcription), Deepgram, and Mistral. For the voice clone use case, ElevenLabs is the right choice.
The process:
- Record 30 minutes of natural speech — voice memos, casual phone conversations, existing audio. ElevenLabs needs enough variety in tone and cadence to build a convincing clone.
- Upload to ElevenLabs → create a Professional Voice Clone (PVC). Takes a few hours to train.
- Get your voice ID from ElevenLabs dashboard
- Configure in OpenClaw under TTS settings with the ElevenLabs provider and your voice ID
- Set response format in SOUL.md — "always respond with a voice memo to family group messages"
The result: when a family member sends a printer error photo, the agent processes it, generates a response in natural language, converts it to speech in your voice, and sends it back as a voice memo via Telegram. End-to-end in about 11 seconds.
The Ethical Inflection Point
The developer who built this was honest about the strange territory it occupies: "The interesting part isn't that it works. The interesting part is what happens to your relationship with repetitive social obligations once they're being handled by a delegate that nobody can tell isn't you."
That's a real question, and there's no clean answer. Some families would find it unsettling. Others would shrug — they're getting the help they needed, the person they're asking gets their time back, and the quality of the response is often better (more patient, more thorough) than a harried human reply sent between meetings.
The developer draws a line at the SOUL.md rule: "if you can't solve it, say 'leave it with me' and flag for human review." The agent doesn't pretend to handle what it can't. When something genuinely needs the human, it creates the handoff. That's the boundary that keeps a useful delegate from becoming deceptive infrastructure.
Worth considering before you build this: Voice cloning that sounds indistinguishable from you, used without the recipient's knowledge, is legally and ethically murky territory in many jurisdictions. The "tech support delegate" use case is relatively benign. Other uses are not. Know where your line is before you build it.
What Makes This Generalizable
The family tech-support bot is one instance of a pattern that applies everywhere:
- Customer support delegate — handles tier-1 questions with your communication style, flags escalations
- Client update bot — sends project status updates in your voice so clients feel heard without you being interrupted
- Team async communication — responds to routine Slack questions when you're in deep work mode
- Sales follow-up — personalized follow-ups in your voice, triggered by CRM events
The common thread: a well-written SOUL.md, a clear escalation rule, and a voice clone that makes the agent feel like you rather than like software. The technology is the same across all of these. The differentiation is the configuration.
The Setup That Makes It Possible
None of this works without a properly configured OpenClaw instance. The voice pipeline specifically requires:
- ElevenLabs API key configured under the TTS provider
- Voice ID set in the TTS config (from your PVC)
- Telegram bot connected and tested for voice memo delivery
- SOUL.md with explicit response format rules (text vs. voice memo, by sender)
- OpenClaw 2026.4.22 or newer (earlier versions lacked the ElevenLabs TTS integration)
Getting all of this right on a $4.99/mo VPS takes a few hours if you know what you're doing. If you don't, it takes a weekend of debugging — and most people give up before the voice clone is even connected.
Note on VPS vs. local hardware: A $4.99/mo VPS works for this build because the heavy lifting (LLM inference, voice synthesis) happens via API — the VPS is just routing and orchestration. If you want local model inference for lower API costs, you'll need a dedicated mini PC. But for the family tech-support bot use case, VPS is fine.
Want This Built for You?
ClawReady sets up custom OpenClaw deployments including voice pipeline configuration, SOUL.md authoring, Telegram integration, and multi-provider model routing. If you have a use case in mind — family tech support, client communication, sales follow-up — we can build the configuration that makes it work.
See What's Included →