For most of the last two years an agent's world ended at the terminal. The launches below push it outward: onto the phone in your pocket, onto Android devices rented by the second, into a panel beside your SSH session, into your Mac's notch, and into the room you are actually standing in.
A phone that drives itself
iPhone Use lets an AI model use your iPhone from the iPhone itself, with no Mac at run time apart from one-time pairing, no extra hardware and no jailbreak. It is written in Swift and Rust, needs iOS 26 or later, and each run shows every step with its screenshot followed by the agent's summary. Showing the screenshots is what makes it usable — you can see what it did before it does the next thing.
Android devices rented by the second
Phonebox gives agents cloud Android devices billed by the second. One API call starts a phone and the CLI and SDK wait until it is ready; the agent reads the screen as a short numbered list, then taps by label, types, swipes and opens apps. It can install apps, move files and change location or language, and idle phones park themselves and cost nothing while keeping their apps and signed-in accounts.
An assistant beside every terminal
Hatoba is an open-source desktop SSH client with an AI assistant and end-to-end encrypted sync through your own Cloudflare account. You describe the problem in a panel beside your terminal and the assistant reads the screen and runs commands. Built with Tauri 2 for Windows, Linux and macOS, it keeps the sync encrypted end to end and under your own account rather than the vendor's.
Answering agents from the menu bar
Muster is a free Mac menu bar app for Claude Code, Codex and Cursor. Option-Tab brings your sessions together and Option-Return answers a supported request; the alert unfolds to show the agent's reason for asking, the whole command, the file and size of an edit, or every option of a question. It also lists every session from the last three hours with its tokens and what they would cost at list prices, so you can see which runs are the expensive ones.
Reading the web and Kindle out loud
Narro is a Chrome extension that reads the page you are on aloud in a natural AI voice, highlighting each sentence as it goes. It can use a cloud voice with your own key or a local model that runs entirely in the browser, and it supports the Kindle Cloud Reader. With a local voice there is no account and it works offline after a one-time download, so nothing you read leaves the machine.
No chat box, just a key
Memoe is a voice assistant that lives in your Mac's notch. You hold one key from any app, say what you need, and Moe answers out loud and does the work, using the ChatGPT or Claude plan you already pay for. The absence of a chat box is the design: voice is an input method for the whole machine, not a window you have to open first.
A voice assistant with no GPU
Sonia Assistant is a local voice assistant combining an LLM with speech-to-text and text-to-speech that needs no dedicated GPU and runs in less than 15 GB of RAM entirely on a local computer. It is multilingual and full-duplex, so the microphone stays open while it talks and you can interrupt it. The only external piece is an OpenAI-compatible endpoint, so you can point it at a model you already run.
Live captions that stay on the Mac
Connecting Captions gives you live on-device captions on a Mac, with each sentence appearing when it is ready. Use it for a presentation, a bilingual room, or any conversation that needs a translation. Speech recognition and Apple Translation run on the Mac itself, so voice never leaves the machine unless you turn on a cloud speech model, and there is no account and no telemetry.
Transcribing whatever the computer plays
Sys2txt records your system audio and transcribes it locally with Whisper. It works on demand, recording until you stop and transcribing once, or in live mode, segmenting every N seconds and printing each transcription as it is created. You can use faster-whisper, openai-whisper or whisper.cpp, and it auto-selects whichever is installed rather than failing part-way through.
A to-do list your agent can read
Nagare is a to-do app for iPhone, Mac and your agent built on the idea that calendar events and tasks are both commitments. Both are one kind of item with an optional time, so you can put studying before the dentist appointment in one ordered list instead of reconciling two views. It is the quietest entry here, and the clearest example of an app designed assuming an agent will read it.
Why this cluster matters
Each of these moves the agent closer to where the work is, and each pays for it with a new trust question: a phone that can book a ride, an assistant that can hear the room, a client that can run commands on a server. The launches worth watching are the ones that answer that by showing you what happened — screenshots per step, a diff per change, captions that never leave the machine.
