用旧 Android 加 6 美元 ESP32 冒充蓝牙耳机，把电话接进 Codex CLI 实时语音

raw POC, filmed on a real call. the voice is codex cli's realtime voice, signed in with a chatgpt plan: no api key, no twilio/sip. it answers the phone's own number. all i built is the plumbing that gets a call's audio to it.

the trick: android won't let any app touch a call's audio, but it hands every call to a bluetooth headset. so a ₹599 ($6) esp32 pretends to be one (hands-free profile, 16 kHz) and passes the audio back to the phone over usb; an app there streams it to the voice.

what broke on the way:

- bluetooth and wifi share the esp32's one radio. audio over its wifi lost ~45% of the voice; over the usb cable, none.

- codex's voice is made to hand every request to the agent behind it (makes sense for coding). on a phone call that meant seconds of silence and the whole menu read out. so on calls the voice gets the shop's handbook (menu, timings, rules) up front and answers itself.

- now it starts ~2 s after the caller stops; the esp32's share is ~75 ms, the rest is network + model. talk over it and it stops mid-word.

in the video i call as a customer and it picks up as "arjun" from a sweet shop in hyderabad: hindi, english, telugu, whichever i speak. i said their sweets are rotten and asked for chicken; it stayed polite and stuck to the shop's rules. on other calls we asked for the day's news and a film's reviews; it went and got them mid-call.

the brain runs on wrkr (am building it): ur chatgpt gets its own linux computer in the cloud, always on. the shop's handbook lives there, so the line never sleeps.

credit to openai: i built the phone setup, but the real-time voice in codex cli is what made it possible. without that, this doesn't exist.

ask me anything on the esp32/android side.