Conduck

Conduck

Your AI, your keys, your voice

0 ratings
Free

Rating summary

Details

  • Released
  • Updated

Features

Conduck screenshot #1 for iPhone
Conduck screenshot #2 for iPhone
Conduck screenshot #3 for iPhone
Conduck screenshot #4 for iPhone
Conduck screenshot #5 for iPhone
Conduck screenshot #6 for iPhone
Conduck screenshot #7 for iPhone
iphone
ipad
watch
🖼️Get Icon
Icons↘︎

About

Conduck is the native Apple client for your self-hosted or BYO-key AI - in your pocket, in the car, and on your wrist. Connect an OpenAI-compatible gateway like OpenClaw, Hermes, Ollama, or LM Studio, or bring an OpenRouter key for a hosted model in about a minute. Your keys, your AI. No Conduck account, no intermediary servers, no relay. Conduck isn't another AI service. There's no model inside it and no intermediary server in the middle. You connect it to an AI you control, then talk or type to it from all your Apple devices. Ask from anywhere • iPhone & iPad: start a chat and type, or hold to talk • Action Button, Shortcuts & Control Center: fire off a quick question without opening the app • Apple Watch: ask straight from your wrist • CarPlay: a hands-free voice conversation on the road • Mac: a full desktop app and menu-bar companion, with a global hotkey and Screenshot & Ask Private by architecture, not by promise • Conduck has no intermediary servers. No accounts, no analytics, no ads, no tracking. • By default, your speech is turned into text right on your device by Apple. Your voice never leaves your device. • Prefer a cloud transcription service? Your audio goes straight from your device to that provider with your own key. We're never in the middle. • Your conversations stay on your device and sync through your own private iCloud. We can't read them. • Your keys and access tokens live in the Keychain, and every connection over the internet is HTTPS. A server on your own network can also be reached without it, if you choose that - and Conduck says so on screen when you do. See what you actually use • Settings keeps a private count of your turns and conversations over time, typed against spoken, for the last week, month, or 90 days. • Broken down by gateway, by device, and by the model each request asked for, with token totals exactly as your gateway reports them. • How often a request lands on the first try, plus the failures that did not, each with the reason it gave. • Measured on your device from your own conversations. Visible only to you, reported to no one. Free to download and use Conduck is free to download and use. No ads, no tracking, no Conduck account. Your AI provider or hosted model may charge separately. What you'll need Conduck is the voice and the interface, not the AI itself. You bring the AI, whichever way suits you: • Start in about a minute: connect a hosted model like OpenRouter with your own API key. Nothing to install, no server to run. • Run your own: point Conduck at a personal AI server you host (such as OpenClaw or Hermes) or any OpenAI-compatible service (Ollama, LM Studio, and the like). It's your server to run and maintain. Conduck simply talks to it. Either way, Conduck walks you through connecting, but it won't create an account or host a model for you. Conduck is for people who want AI on their own terms: their model, their keys, their data, in a fast, native app built to stay out of the middle.
Show more

What's New in Conduck

1.5

August 28, 2026

Usage arrives in Settings: every turn Conduck sends, counted on your device and seen by nobody else. Coming from 1.2? Everything from 1.4 and 1.3 is in here too. Usage • How much you actually ask: turns, conversations, and typed against spoken, over 7, 30 or 90 days, or all of it. • Charted by turns, tokens, models, devices or gateways, and broken down by gateway, by device (iPhone, iPad, Mac, Apple Watch, CarPlay) and by the model each request asked for. • Token totals exactly as your gateway reports them, with cached input and reasoning output called out separately, so a bill that looks wrong has something to check against. • Reliability at a glance — how often a first try lands, what a retry rescues — plus the individual failures and the reason each one gave. • Your heaviest conversations, ranked, so it is obvious which thread ran up the numbers. • Where your gateway reports nothing, the numbers say so rather than counting it as zero, and every figure says how much of the range it actually covers. • The records outlive the chats they describe, so tidying your conversation list no longer punches a hole in the trend. Clear usage history removes them everywhere, once. • Counted on your device from your own conversations. No account, no analytics, nothing reported to anyone. CarPlay • A session that dies while it is starting can no longer leave the car's microphone wedged. • A reply you have already heard in the car no longer sits there unread. • A voice session that ends in silence signs off out loud instead of just stopping. Sync and Watch • Forgetting a gateway, or deleting a voice endpoint, on one device now lands on the others even if they were asleep when you did it. • Apple Watch no longer flashes busy UI on its way into a capture. Also in 1.4 • Point Conduck at Ollama, Open WebUI, LM Studio or any OpenAI-compatible server on your own network by its IP address or its .local name, with no certificate to set up first. • The address field says what you are trading: the connection is not encrypted, and it works only while you are on that network. Everything beyond your own network still requires HTTPS. • A dictated note on Apple Watch no longer waits forever for its conversation, and can never land in one you did not ask for. Also in 1.3 • Files: every conversation gets its own output folder on your file server, so a reply offers only what is really in that folder, and nothing is fetched until you tap it. Returned files keep their real names, and cloning a conversation brings its files with it. • Your conversation list: rows show working, answered, failed or idle at a glance, and unread and failure marks belong to your account, so reading on one device clears them on the others. A new chat starts on the gateway your last one used. • Privacy and trust: a gateway that redirects to another host, or silently downgrades an encrypted connection, is refused. Replies no longer load images from the internet. • Speed: long replies render far faster, and a reply big enough to take the app down is now bounded. • When something goes wrong: a failed turn keeps its reason across a relaunch, a photo your model would not accept no longer poisons the rest of the chat, and every error names the AI you configured. • Voice: a cloud voice that cannot play falls back to the Apple voice and says so, instead of ending the turn in silence. • Mac and iPad: every control properly clickable, a sidebar that fits the window, drop a file anywhere on the conversation, and a notification when a reply lands. • Plus Copy conversation, an Open Source Licenses screen, and a long tail of copy and layout fixes.

More