Self-hosted outbound telephone calling for AI agents: a SIMCom SIM7600G-H modem places a real mobile call, and the xAI Realtime voice API holds the conversation. See docs/TELAIUS_CONCEPT.md for the vision, docs/HARDWARE_NOTES.md for what is proven on real hardware, and docs/VOLTE_NOTES.md for IMS/VoLTE findings.
Status: M0 (USB call + two-way audio) and M1 (live AI conversation) work as CLI prototypes. M2 (MCP interface) is not started.
Requires Python 3.11+. Linux uses the kernel's option and qmi_wwan devices; macOS uses the
raw libusb fallback (brew install libusb). adb is also needed to install the optional 16 kHz relay.
python3 -m venv .venv
.venv/bin/pip install -e .
cp .env.example .env # gitignored; fill in XAI_API_KEY, optionally TELAIUS_VOICEOn Linux, Telaius automatically uses the stable SimTech links for USB interfaces 2 and 4:
/dev/serial/by-id/*SimTech*-if02-port0 # AT commands
/dev/serial/by-id/*SimTech*-if04-port0 # call PCM
/dev/cdc-wdm0 # QMI
The user must belong to dialout. Install the included udev rule to grant dialout access to
cdc-wdm and keep ModemManager away from this modem, then unplug/replug it:
sudo usermod -aG dialout "$USER"
sudo install -m 0644 udev/99-telaius.rules /etc/udev/rules.d/
sudo udevadm control --reload-rules
# log out/in if dialout was just added, then reconnect the modemIf ModemManager is already managing the modem, stop it before the first reconnect
(sudo systemctl disable --now ModemManager) or rely on the installed ignore rule. Set
TELAIUS_TRANSPORT=tty to require kernel devices or TELAIUS_TRANSPORT=usb to force the raw
libusb fallback. Automatic mode prefers kernel devices when they are present.
# AT console (if2 of the modem)
.venv/bin/python -m telaius.at ATI AT+CPIN? AT+CREG? AT+CPSI?
# QMI (if5): IMS registration state and carrier MBN profiles
.venv/bin/python -m telaius.qmi status
.venv/bin/python -m telaius.qmi mbn-list
# M0: call, play a WAV into the call, record the far end
.venv/bin/python -m telaius.calltest +48XXXXXXXXX phrase.wav recordings/rx.wav
# M1: let the AI talk on a real call (ru or pl prompt)
.venv/bin/python -m telaius.voicebridge --no-call # check the xAI session only
.venv/bin/python -m telaius.voicebridge +48XXXXXXXXX --lang pl --voice rex --max-seconds 120
# 16 kHz VoLTE audio via the relay on the module (one-time: AT+CUSBADB=1, reset, then install).
# relay/telaius-relay is prebuilt (Git LFS); rebuild with: docker build -t telaius-armbuild relay/ && (cd relay && docker run --rm -v "$PWD":/src telaius-armbuild make)
relay/install.sh
.venv/bin/python -m telaius.voicebridge +48XXXXXXXXX --audio relay
# a call with a goal: the bot introduces itself, pursues the task, then reports
.venv/bin/python -m telaius.voicebridge +48XXXXXXXXX --audio relay \
--task "Записаться на стрижку в субботу после 12:00" \
--callee "барбершоп Brzytwa" --context "Клиент: Алексей, обычная мужская стрижка"Each call writes calls/<date>_<time>_<callee>/ (gitignored) with call.mp3 (both sides, via ffmpeg; call.wav if ffmpeg is missing), a readable transcript.txt and result.json: outcome (done / partial / failed / no_task / unknown), summary for the owner, end reason (bot_goodbye, remote_hangup, max_duration, not_answered: …), network and the full transcript. The bot reports through its end_call tool. If the callee hangs up first, the bot is asked for the result afterwards (report_result).
Prompts are config files in prompts/: ru.toml / pl.toml hold each language's texts (free chat, intros incl. intro_formal, the [task] template and rules, recording facts, voice tags, gender lines), settings.toml holds the owner's name, the bot name per voice gender (Telaius / Телайя), voice genders, transcription hints and tool descriptions. Adding prompts/<lang>.toml adds a language to --lang. TELAIUS_PROMPTS_DIR points to another folder; --prompt-file replaces the prompt for one call.
Calls go to calls/, and test audio (calltest, --no-call) goes to recordings/; both are gitignored. Only call numbers whose owners agreed to be called by an AI assistant.