Installation
Everything except Asterisk and the ATA runs on your Mac.
1. Clone and run the setup script
git clone https://github.com/xvilo/claude-code-deskphone.git
cd claude-code-deskphone
scripts/setup.sh
scripts/setup.sh:
- installs
whisper-cppandffmpegwith Homebrew, if missing - runs
npm install - downloads the multilingual whisper
smallmodel (~470 MB) tomodels/ - registers this clone as a local plugin marketplace and installs the
deskphoneplugin into Claude Code
WHISPER_MODEL=base scripts/setup.sh gets a smaller, faster but less accurate model. Then point DESKPHONE_WHISPER_MODEL at models/ggml-base.bin.
The plugin is installed from your clone rather than straight from GitHub on purpose: its channel server needs the clone’s node_modules, and the daemon runs from the clone anyway.
2. Create the config files
cp .env.example .env # read by the daemon
cp asterisk/.env.example asterisk/.env # read by Asterisk
Generate two passwords and put them in asterisk/.env:
openssl rand -hex 12 # PHONE_PASSWORD: the ATA's SIP password
openssl rand -hex 16 # ARI_PASSWORD: the daemon's Asterisk password
Copy ARI_PASSWORD into .env as DESKPHONE_ARI_PASSWORD. You set DESKPHONE_ARI_URL in the next step, once you know where Asterisk runs.
3. Set a PIN
npm run set-pin
While your Mac is locked, the phone asks for this PIN before it accepts prompts, approvals or answers. See Usage.
4. Pick a voice (optional)
The daemon speaks with the macOS system voice unless DESKPHONE_VOICE is set. say -v can’t select Siri voices by name, but if you choose a Siri voice under System Settings → Accessibility → Spoken Content → System voice, the daemon uses it. You can also download a Premium voice there (for example Zoe (Premium)) and set DESKPHONE_VOICE="Zoe (Premium)" in .env.
Each reply is spoken in its own language. The daemon detects it with macOS’s language recognizer. When the reply isn’t in DESKPHONE_VOICE_LANGUAGE (default en), the daemon picks an installed voice for that language. It prefers the home locale (nl_NL over nl_BE) and Premium or Enhanced voices, so downloading one, for example Xander (Enhanced), is enough. To choose one yourself, set DESKPHONE_VOICE_NL="Xander (Enhanced)". To use a Siri voice for your main language, make it the system voice and set DESKPHONE_VOICE_LANGUAGE to that language.
Short replies such as “Ja.” or “OK, done.” are hard to recognize. When the recognizer isn’t sure, the daemon keeps to a language already in play: the session’s last clearly detected reply language, then DESKPHONE_LANGUAGE, then any language with a DESKPHONE_VOICE_<LANG> setting, then DESKPHONE_VOICE_LANGUAGE. If you often get short replies in a second language, setting DESKPHONE_VOICE_<LANG> for it helps from the first reply on.
The words the daemon adds around a reply follow the reply’s language too: “(code on screen)”, “a link to …”, “the rest is on screen”, the “Extension 101” and “Recap” or “Last time” prefixes, and the instructions after a question from Claude. They exist in English and Dutch; other languages get the English wording. Menu prompts such as the PIN, the 100# menu and the voicemail announcements stay in English.
Next: run Asterisk.