softphone

Go SIP user agent for the cia.mn softphone, built on diago / sipgo. It registers to the SIP provider (ip-phone.mobinet.mn), places and answers calls, and forwards inbound callers via DTMF.

End goal: a browser softphone — the browser does WebRTC audio and this service bridges it to SIP. Progress:

1. Register to the SIP registrar — ✅
2. Place / answer calls (G.711 audio), DTMF, call forwarding — ✅
3. Browser front-end over WebRTC, bridged to the SIP leg — next

Configure

cp .env.example .env   # fill in SIP_USER / SIP_PASS (from Mobinet)

All config is environment variables (loaded from .env if present); see .env.example for the full list.

Run

go run .                 # register and stay online (answers inbound calls)
go run . call 99xxxxxx   # register, then place a call and play a test clip

• Inbound (while running): incoming calls are answered. With FORWARD_TO set, the caller is bridged to that number when they press 1; otherwise it plays a prompt and echoes their audio. At most FORWARD_CONCURRENCY forwards run at once — extra callers are held on hold in a FIFO queue up to QUEUE_TIMEOUT.
• Outbound (call <number>): dials and plays a test clip on answer.

Custom audio (prompt / hold music)

Calls use 8 kHz mono 16-bit WAV. Convert any MP3 with ffmpeg:

ffmpeg -i sounds/start.mp3 -ar 8000 -ac 1 -c:a pcm_s16le sounds/start.wav

The forward flow uses three clips (defaults shown):

PROMPT_FILE=sounds/start.wav                # greeting while waiting for "press 1"
CONNECTING_FILE=sounds/calling-forward.wav  # "connecting…" while the forward leg rings
HOLD_FILE=sounds/waiting-queue.wav          # hold music while queued (slots full)

So a forwarded call sounds like: greeting → press 1 → "connecting…" → far end answers → talking; if all forward slots are busy, the caller hears the hold clip until one frees.

A path that doesn't exist falls back to a bundled demo clip. In Docker the compose file mounts ./sounds → /sounds, and the container's working dir is /, so the default sounds/*.wav paths resolve to /sounds/*.wav. For plain docker run, add -v "$PWD/sounds:/sounds".

Deploy (Docker / GHCR)

⚠️ Inbound calls need a publicly reachable host. Behind NAT the carrier can't reliably deliver the call's ACK/media. Run on a host with a public IP. BIND_HOST is the local interface to bind (leave empty to auto-detect — don't put a 1:1-NAT public IP there); PUBLIC_HOST is the public IP to advertise. (Outbound works from anywhere.)

The GitHub Actions workflow (.github/workflows/docker-publish.yml) publishes a multi-arch image to GHCR on every push to main and on v* tags. Pull and run it on your server:

docker run -d --name softphone \
  --network host \
  --env-file .env \
  -e PUBLIC_HOST=<public-ip-of-this-host> \
  ghcr.io/cia-mn/softphone:latest

(If the public IP is assigned directly to the host's interface — ip addr shows it — you can instead set BIND_HOST to it and skip PUBLIC_HOST.)

• --network host lets SIP + the dynamic RTP ports work without mapping a range.
• Config comes from --env-file .env or -e flags — no secrets are baked into the image.

Build locally instead (needs Docker BuildKit/buildx):

docker build -t softphone .
docker run --rm --network host --env-file .env -e PUBLIC_HOST=<public-ip> softphone

Ports to open (firewall / security group)

network_mode: host binds directly to the host, so there's no Docker port mapping — open these inbound UDP rules on the server (cloud security group and/or host firewall). Pin fixed ports so they're predictable:

Port (UDP)	Purpose
BIND_PORT — e.g. 5060	SIP signalling
RTP_PORT_START–RTP_PORT_END — e.g. 10000–10100	RTP/RTCP audio

Recommended server .env for deployment:

BIND_HOST=                 # empty = auto-detect this host's interface IP
PUBLIC_HOST=<public-ip>    # the IP advertised to the carrier (Contact/SDP)
BIND_PORT=5060
RTP_PORT_START=10000
RTP_PORT_END=10100

(If ip addr shows the public IP directly on the interface, set BIND_HOST to it and leave PUBLIC_HOST empty — then the SIP Contact is public too.)

Each call leg uses 2 ports (RTP + RTCP); a forwarded call has 2 legs (4 ports), so a 100-port range handles plenty of concurrent calls. Everything is UDP (only SIP_TRANSPORT=tls/tcp would add a TCP port). Outbound has no inbound-port requirement — only inbound calls/media do.

Troubleshooting

• Raw SIP messages: set LOG_LEVEL=debug.
• 403 after auth → wrong SIP_USER / SIP_PASS, or the username isn't the bare local number (Mobinet uses e.g. 75151407, not +976…).
• Inbound No ACK received or one-way audio → NAT. Deploy on a public-IP host (see above) or port-forward a fixed UDP port.