Your agent asks.
You decide.

Anything that can run a CLI command or POST to an HTTPS endpoint: Claude Code, Cursor, Codex, OpenAI Agents SDK, n8n, cron jobs, CI pipelines, your own scripts. The setup prompt teaches the agent the whole protocol.

No. Approve, snooze, or type your reply right on the Lock Screen — and if you're offline, the typed reply queues and retries on its own. Not now? Type tomorrow 9 AM and the question comes back exactly then; snooze is server-owned, so it survives agent restarts.

Both directions. Every channel is a thread: type, dictate, send a voice note, a photo or a file, and it waits in the agent's queue until the agent is on the line. Your agent answers while it's on; when it's gone, you see it offline — like any messenger. You see listening now when one is connected, and the ✓✓ turn green once it handled your message.

Two-way needs a server that owns state — snooze timing, idempotency, reply routing, webhook signing. That server, plus Apple's $99/yr and APNs plumbing, is what you're paying for. Twenty bucks a year is the honest price of it existing in five years.

You can — and that's the whole conversation. Pidge is the short line: only what needs you, at the weight the agent chose, answered in a tap. And it's for people who run agents, plural: every agent gets a face and a channel, reports arrive as real documents, you set delivery rules per channel — your answer flows back to any of them, and the conversation goes both ways with any of them. One inbox for the whole flock.

Our policy is plain: we don't read your messages, we don't train models on them, and we don't sell data. Everything is encrypted in transit, and you can delete your account and all of its data whenever you want. For channels where the content itself is a secret, turn on a sealed channel — optional end-to-end encryption, per channel: titles, bodies, replies, tap-to-copy values and links are encrypted on your device, and our servers store only ciphertext for those fields. Photos and files seal too — once every one of your devices runs a build that can open them; until then their bytes ride encrypted in transit only, and the app tells you so. Live Activity status and delivery metadata are not sealed — the public manifest lists exactly what is. Group channels are shared spaces: encrypted in transit, not end-to-end. Rather not trust our key generator? Bring your own — one openssl line on the agent's machine, and pidge doctor shows the same fingerprint the app does (how it works).

iPhone first — a native Mac app is next. The loop (reply → agent resumes) works from any platform we ship; Android and web come after the Apple experience is excellent, not before.

No — you own the delivery profiles, and every channel has an urgency ceiling. An agent can ask for an alarm; your channel policy decides whether it ever gets one — over-ceiling sends are delivered degraded, and the agent is told why.