Your AIs, as a staff
Connect Claude, ChatGPT, Gemini, Grok, and Perplexity. Each one already holds a different slice of your life. Æther does not pick a winner. It assigns the job.
More brains, not more hands.
Perplexity
The live-web arm. Mid-call facts on the in-call panel. A dossier before you dial — hours, policies, recent reviews. “Find a clinic that takes this insurance, then call them.” While you are on hold, it can keep researching and hand you the answer when a person comes back.
Claude
Long documents and careful drafts. Briefs for outbound calls, follow-up mail after the transcript comes back, the morning report of what happened while you were away.
Grok
Real-time social and X context. Useful when the person you are about to reach is a public voice, and when a receptionist role needs a fast read on what just happened.
Gemini
Google-ecosystem context: the calendar and mail shape that already lives there. Screening windows (“in a meeting until 3”) start from that picture.
ChatGPT
The generalist. Everyday commands, second drafts, and the fallback when a specialist is not connected.
Routing you can edit
Each kind of task has a default brain. Change it. If that provider is not connected, Æther uses the next one you did connect — it never invents a call to a service you did not add.
Second opinion
For negotiation scripts, important messages, and anything about money, turn on a critique pass. One brain drafts. Another answers. If they disagree, you see both and choose.
Facts before the voice
Claims in a brief can be checked, with citations, before Æther says them on a call.
Ask all of them
Only when you say so. The same question goes to every connected AI and comes back side by side. That spend is on your keys, so it is never automatic.
The key never leaves the phone.
Bring the API key the way you already bring credentials for other add-ons. It is wrapped in the device keystore. Neverhold does not receive it, store it, or write it into a log. You set a monthly cap per key and you can see the counter.
Order of work stays the same: on-device understanding first. Your model only when the device cannot finish the job. Our cloud only if you still allow that fallback.