Android

codai for Android uses your phone the way you would: it sees the screen as a list of buttons and texts (through an Accessibility service), taps, types, scrolls and swipes — and stops for an ask card before anything risky. Android 11 or newer. Apache-2.0, source at github.com/codai-ro/codai-phone.

Install

Pair it with your account

The first screen after the legal consent asks how codai should think:

• Sign in with your codai account — email + password, or Sign in with Google, natively in the app (no browser). This mints a codai_ key for this phone, labelled Android · <model> in the hub; signing out revokes it.
• Use your own AI key instead — skip sign-in, then Settings → Providers: OpenAI, Anthropic, Google Gemini, OpenRouter, GitHub Copilot or any OpenAI-compatible endpoint. Keys stay in Android's encrypted preferences and go only to that provider.

Your phone appears in hub.codai.ro → Security → Devices; revoke it there if it is lost.

Permissions

The onboarding wizard walks through these in order. Only Accessibility is required for phone control; everything else is optional and can be switched on later from Settings → Permissions.

Android shows a warning that the accessibility service can "view and control the screen". That is accurate. Password fields are masked by Android, but other on-screen text — a bank balance, a 2FA code — is visible while a task runs; that is why banking, wallet, payment, authenticator, password-manager and system Settings apps are treated as sensitive and require approval in the default autonomy level.

What the agent can do on the phone

• Use other apps — open, read, tap, type, scroll, swipe. "Open WhatsApp and tell me the last message from Ana."
• Answer questions about the phone — notifications, alarms, contacts, calendar, recent SMS, foreground app, Bluetooth, battery, storage.
• Small chores — set alarms and timers, add calendar events, open a link, share text, copy to the clipboard, reply to a notification, start a call.
• Files — list, search, read, edit and move files in your storage (APK build).
• Terminal — a real Linux-style shell through the companion app or Termux, optionally jailed to a scratch home (Sandbox switch in the agent-mode sheet).
• Work offline for small things — optional on-device model (Gemma 4 E2B, ~2.6 GB, or the fine-tuned codai-nano) under Settings → Models. Modes: off (default), fast-lane (mechanical follow-up taps run locally), local-first (every step starts on the phone and hands over to the cloud when stuck). Best on recent flagships; 4 GB phones should stay off.
• Ask before risky actions. Sending a message, paying, or acting inside a sensitive app pops an ask card. Messaging apps have their own switch: ask (default), block, allow.
• Several tasks at once — one session per window (split-screen / pop-up on One UI); tasks that need the screen queue politely.
• Keep a plan and notes — a live todo checklist under the chat header and a local Memory browser (/memories/<file> survives sessions; nothing leaves the phone).

Autonomy

Settings → Agent → Autonomy, also in the shield chip next to the composer:

Stop in the chat or on the bubble ends a task at once; every turn also has a time budget (4 minutes by default, Settings → Agent).

Drive your desktop from the phone

With codai desktop open on a computer signed in with the same account and Serve as a remote host for my phone switched on, the phone can:

• run terminal commands and read/write files on that computer (terminal tool with the relay backend, inside the desktop's approved folders);
• dispatch a whole task to the desktop — it is picked up within about 30 seconds and executed there while your phone shows the live transcript with a REMOTE badge;
• take the reverse hand-off: Continue on phone from the desktop sends the conversation here and wakes the app by push (builds with Firebase configured).

Turn Session sync on (Settings → Advanced) to see the same conversation on every device. The device doing the work is the executor; the others are viewer (watch) or editor (send, answer, cancel). Protocol details: Shared sessions.

Billing

The app is free. What costs money is the model that does the thinking:

• codai account — usage is billed by your plan (tasks per month, then wallet or overage). Inside the app, Cost · Receipt shows the server's receipt for a session, and Settings lets you set a per-session budget so a task cannot run away.
• Your own provider key — that provider bills you directly; codai takes nothing.
• On-device steps cost nothing.

Where you buy depends on the build:

Plans, prices and what a "task" is are on the limits & pricing page.

Privacy

Stays on the phone: your keys and sign-in (encrypted preferences), the screen contents codai reads unless a step needs the cloud model, everything the on-device model handles, chat history (local database), memory notes, and per-turn trace files if you enabled them for debugging (Download/codai-nano/, never uploaded).

Leaves the phone: your prompt, the screen summary and tool results — to the AI backend you configured (codai gateway by default, or your own provider). With Session sync on, the step-by-step transcript is mirrored to the gateway so your other devices can follow. Builds with Firebase send the push token used to wake the phone for dispatch.

Never: no analytics SDK, no ad SDK, no crash reporter. No message sent, payment made or setting changed in a sensitive app without an ask card you tapped.

Training on your conversations is opt-in (default off) — Settings → About → Training consent, or PATCH /v1/account.