AI telephony

LynBro PBX: an AI phone agent on our own Asterisk, from prototype to public signup

An AI agent that answers calls in Danish, English and Ukrainian, transfers to a person and learns from unanswered questions. Public since 21.10.2026.

The problem

Small businesses miss calls while they work: on a roof, at a treatment chair, in a meeting. Voicemail is rarely played back, and a human receptionist costs money for every call. We wanted an AI agent that answers in the caller's language, says plainly that it is an AI, takes a proper message or books a time, and passes the call to a person when that is what the caller needs.

We set hard limits from day one. The only paid dependencies allowed were the server, the SIP operator, the domain and the model providers, OpenAI and Google; no hosted voice-agent platforms. The first technical route then failed on the one feature that matters most. When the operator forwarded calls straight to OpenAI's SIP endpoint the audio was clean, but a transfer to a human was impossible: the operator does not support SIP REFER, and the handover never arrived. A home-made SIP bridge in Python before that had produced choppy audio. We had to own the call itself, not just the conversation.

What we built

On 18 September 2026 we moved the call path to our own Asterisk 22 with FreePBX 17. Asterisk streams the call to a Python sidecar over WebSocket in G.711 µ-law, the same format OpenAI Realtime accepts, so no audio is resampled on the way. Interruptions are handled by flushing Asterisk's own playout buffer, which means the agent stops talking the moment the caller starts. When the AI part of the call ends, the sidecar returns a decision and the dialplan rings the right person or queue, so the transfer happens inside our own telephony.

The control plane at pbx-app.lynbro.dk is a separate FastAPI and MySQL application with a small CRM, a knowledge base in Meilisearch, call logs and settings for each customer. The greeting is fixed text rather than model output. It always says that the caller is talking to an AI and that the conversation is written down as text, not recorded; raw audio is never stored.

Questions the agent cannot answer are logged as gaps and appear under Questions in the panel. Someone on the customer's team answers in a single field, a model turns the answer into one to five short facts, and the next caller gets the answer. A caller simulator with a scripted persona and a judging model replays scenarios for every line on a weekly cycle, so changes are tried before real callers hear them. The agent can also look up and book times in bkg.dk, our booking system.

Self-signup opened on 21 October 2026 with a CVR check, email confirmation and an SMS code. Outbound calls are limited to Danish numbers, with a daily cap per plan.

Facts and results

21.10.2026public site, pricing and self-signup went live
0 resamplingµ-law all the way from Asterisk to the model
4 plansPrøv: 30 free minutes for 14 days; Start 499, Pro 1,299 and Business 2,999 DKK a month ex. VAT
< 15 scalls shorter than 15 seconds are not billed on the public plans
783automated tests green on main (04.11.2026)
3 languagesDanish, English and Ukrainian in calls and in the panel
1 pilota Danish construction company's line answered by the agent since 28.10.2026

Stack

Asterisk 22.10 and FreePBX 17 on a dedicated telephony server; chan_websocket media in G.711 µ-law at 8 kHz with 20 ms frames; a Python asyncio sidecar; OpenAI Realtime (gpt-realtime-2.1) for speech-to-speech; FastAPI and MySQL for the control plane; Meilisearch for the knowledge base; Stripe for subscriptions; a panel in Danish, English and Ukrainian. The two servers are 0.9 ms apart, and the control plane is 3.5 ms from the OpenAI API.

What we learned

Transfer to a human looked finished for two weeks and never worked. Our test probes passed because they never reached the transfer step. On 14 October a real call showed that the sidecar dropped its leg too early, so the dialplan received hang-up instead of transfer. Now a transfer only counts as tested when the decision line appears in the Asterisk log. Real calls taught us more: early on the agent took background noise for a yes and saved a wrong surname, so names are now read back and confirmed before saving. A booking call that timed out after 8 seconds told a caller it had failed while the server completed it; write actions now wait up to 26 seconds. And when the OpenAI key behind a line ran out of quota, callers heard silence. Now they hear a fallback message, and the business gets an urgent callback task.

Need something like this for your business?

Describe the task in the form. After a free scoping talk we send a plan and a quote, and we tell you which parts of these products we can reuse for you.

Get Started Today