The Pitch

claude-phone by NetworkChuck gives your Claude Code an actual phone number. SIP, 3CX, ElevenLabs for the voice. You call an extension and talk to Claude. Claude can call you back with alerts. "I've finished the migration, want me to push?"

The idea is brilliant. Talk to your coding agent like you'd talk to a team member over the phone.

The Reality

I've tried the lot. claude-phone, Telnyx with OpenClaw, Discord voice channels. The fundamental problem is the same everywhere: the STT-to-model-to-TTS loop, plus all the intermediate network hops, makes natural conversation impossible. You say something, wait two to four seconds, get a response, say something else. It's like talking to someone on a satellite phone in 1997.

The latency breakdown: your speech gets transcribed (STT, ~500ms), sent to the model (network hop + inference, 1-3s), the response gets synthesised (TTS, ~500ms), and streamed back. That's 2-4 seconds minimum per exchange. Humans expect conversational turn-taking under 200ms. The gap is an order of magnitude too wide.

When It's Worth It

Demo day. Investor meeting. Pub bragging rights. Getting a phone call from your coding agent saying "the refactor is done, all tests passing, want me to open the PR?" makes people's eyes go wide. As a party trick it's unbeatable.

Can't see it being used organically in a day-to-day workflow though. For actual voice-driven coding, speech-to-text tools that type into your terminal are faster and simpler. The phone call adds latency and complexity without adding capability.

Keep it in the back pocket. It's impressive when you need impressive.