Paseo vs Pocketd

Run Claude Code, Codex, and OpenCode on your machine from your phone. Paseo is a self-hosted app that connects to your local dev environment so your agents run with full access to your actual setup - not a cloud sandbox. Key features: - Voice support: talk to your agents hands-free with on-device speech processing - Multi-provider: use Claude Code, Codex, or OpenCode from one interface - Git worktrees: agents work in isolated branches, review diffs and ship from the app - Real-time streaming: live agent output and push notifications when tasks complete - Self-hosted: runs on your laptop, home server, or VPS. Your code stays on your machine - Remote access: optional end-to-end encrypted relay for connecting from anywhere - Open source and free Requires a running Paseo server on your machine. See paseo.sh/docs for setup.

Every other offline AI app puts a stranger in your pocket. This one puts an assistant. Ask "what's on tomorrow?" and it reads your calendar and tells you. Ask it to remind you to take the bins out at 2am and it files the reminder. Ask for one every weekday at 7 and it sets one that repeats. Ask how you slept and it reads your Health data and answers. All of it happens on the phone. No account, no server, and no request leaves the device to answer any of it. WHAT IT DOES FOR YOU • Calendar — what's on today, whether Thursday is free, when your next meeting is. And it adds events when you ask. • Reminders — what's overdue, what's due today. Set new ones, including repeating ones, and tick them off when they're done. • Health — how you slept, how active you've been, whether your resting heart rate has moved. • Scheduled tasks — have it read your day at 7am and tell you what needs preparing, every morning, while you're still asleep. Each one is a switch you turn on yourself. All three start off. PRIVACY YOU CAN CHECK, NOT JUST READ ABOUT Most apps promise this in a paragraph. Here it is a property of how the app is built: • Requests arriving over the network can never reach your calendar, reminders or Health data, and can never change them. That is decided by the connection the request came in on, not by anything the caller claims about itself. • A scheduled task running at 7am can read but never write. Nobody is watching at 7am, and a meeting invite is written by a stranger. • Nothing is ever deleted. The assistant can add a reminder and tick one off — both undoable — and cannot destroy anything. • Settings › Privacy lists every permission, every byte the app stores, and the only two addresses it ever contacts, with a real delete button. The whole source is on GitHub, MIT licensed. You can read the gate yourself. RUNS REAL MODELS Qwen 3, Llama 3.2, Gemma 3, SmolLM2 and anything else in GGUF — search Hugging Face from inside the app. It tells you before you download whether a model actually fits your iPhone's memory, instead of letting you find out when it crashes. Vision models too: send it a photo and ask about it. ALSO: YOUR PHONE IS AN AI SERVER Press Start and every device on your Wi-Fi can use the model through the two standard local chat APIs your tools already speak. Point any client at the phone's address on port 11434 and change nothing else. The phone serves its own chat page, so a laptop needs nothing installed. WORTH KNOWING FIRST • Serving to other devices needs the app on screen — iOS closes the socket when an app is backgrounded. • One generation at a time. A phone is not a datacentre. • The phone gets warm and will pause rather than cook itself. • iPhone with iOS 18 or later. A model needs 2–4GB of free memory; the app tells you before you download. Free. No subscription, no account, and no usage reporting you cannot switch off.