Phones are still where real business happens. Missed calls are missed customers, and most small teams can't afford a 24/7 receptionist. So I've been building Tel-Agent — an open-source AI phone assistant: an agent that answers your business line, talks naturally with callers, and takes action — booking appointments, answering questions, taking messages — then hands you a clean summary.
🔗 Website: tel-agent.com
⭐ GitHub repo: github.com/Dpro-at/Tel-Agent
What it actually does
- Answers every call, 24/7. No voicemail black hole. The agent picks up, greets the caller in a natural voice, and handles the conversation.
- Understands intent, not keywords. It's built on modern LLMs, so callers can speak normally — no "press 1 for sales" menus.
- Takes real actions. Booking a slot, capturing a lead, answering FAQs from your own knowledge base, or escalating to a human when it should.
- Reports back. Every call ends with a transcript and a summary, so you know exactly what happened while you were away.
Why open source?
Voice AI is too important to be a black box. Your call data is sensitive — customer names, numbers, intents. With an open codebase:
- You can self-host it and keep call data under your own control.
- You can audit it. No guessing what happens to recordings or transcripts.
- You can extend it. Every business has weird edge cases; open code means you can wire in your own CRM, calendar, or language.
Tel-Agent is released under AGPL, so improvements flow back to the community.
How it works (the short version)
The pipeline is the classic voice-agent triangle, tuned for latency:
- Telephony layer receives the call and streams audio.
- Speech-to-text transcribes the caller in real time.
- The LLM agent decides what to say and which tool to call (calendar, FAQ, message-taking).
- Text-to-speech answers back — fast enough that the conversation feels human.
The hard part isn't any single step — it's keeping the round-trip under ~1 second so the caller never feels like they're talking to a robot on a delay. Interruption handling ("barge-in"), silence detection, and graceful fallbacks took more iteration than everything else combined.
What's next
The roadmap is public and moving: multi-language support (including Arabic — underserved in voice AI and personally important to me), deeper calendar integrations, and better analytics on call outcomes.
If you've ever wanted your business phone to just handle it, or you're curious how real-time voice agents are built — check out the repo, star it if you like the idea, open an issue, or just say hi in the discussions:
👉 github.com/Dpro-at/Tel-Agent · tel-agent.com
Every missed call is a missed opportunity. Now your phone can answer for itself.
Top comments (0)