Aporion gives AI real capabilities over your tasks, devices and apps.
Capabilities
Aporion works the computer it runs on, takes tasks in plain language and answers to the phones and laptops you pair: with your permission, where you can watch it, and within limits you set.
On your computer
Aporion opens and focuses your Windows apps, types and presses keys, then reads the text back through UI Automation to check the result. On the web it drives Edge with a profile of its own.
Tasks
It plans the work, carries it out on a live timeline you can follow, and can only call it done by citing results that show each success criterion.
Activity observation
Off until you turn it on. Then it notes which program is in front, its category and for how long, never what you do in it.
Paired devices
Pair a phone or laptop through the hosted relay at relay.aporion.sh, or one you run yourself, by confirming a matching code on your computer. Both ends dial out, and the relay passes sealed messages it can neither read nor forge.
01 Product
Aporion is one operator for your own computer. Tell it what you want done: it plans, uses your tools and apps, and checks that the work actually happened instead of taking the model’s word for it.
Say what you want done. It reads and observes before it acts, and verifies after.
It opens and focuses windows, types and presses keys, and reads text back through UI Automation to check the result.
It drives your installed Microsoft Edge with a profile of its own, never your everyday one.
Work continues from the tray with the window closed, and on schedules that each pass a supervised test run before they can be turned on.
An account or service counts as connected only after its connection test passes.
A work profile you own, and ranked automation ideas that run once with you watching before they go on a schedule.
02 How it works
Everything a model asks for is a proposal. Classification, policy, approvals, locks and budgets are enforced in ordinary code after the model speaks and before anything happens. There is no second path to a side effect.
The proposed call is checked against the tool’s input schema.
The tool’s own code names the action and its targets. What the model says about its intent is never an input.
Kill switch, grants, autonomy, deny rules, always-ask classes, your rules, defaults: the first layer that matches decides.
Consequential actions wait for you. An approval is bound to a digest of the exact action; if anything changes, it asks again.
Only one task holds your mouse and keyboard at a time, so two agents never fight over them.
The call is recorded as executing before it runs, so a crash can always be reasoned about.
It runs with a time limit and can be stopped mid-flight. Anything interrupted is reported, never silently retried.
Done is never taken on the model’s word. Completion must cite evidence, and the task is marked verified, partially verified or unverified.
03 Control
Direct Aporion from a paired phone or laptop while your computer stays where it is. Your computer dials out to a relay; nothing on it listens for connections from the internet.
The Aporion web app, installable on your phone or laptop. Pairing is confirmed on your computer, with a matching code on both screens.
Routes end-to-end encrypted, signed messages between them. It can delay or drop a message, never read or forge one, and you can host it yourself.
Re-checks every command against that device’s permissions and its own policy. Approving a consequential action from a device takes your PIN.
On your computer, the Devices screen lets you rename a device, limit what it may do, silence its notifications, set when its access ends, or revoke it at once.
Notifications are encrypted for that one device before they leave your computer. They say little unless you choose more detail on that device.
04 Specialists
Name a specialist for a kind of work and give it its own permissions, budget and memory. When the operator hands it a task, permissions only ever narrow.
05 Security model
Remotely reachable but never publicly exposed. Bounded at every step.
Your model key, device keys and connection tokens live in Windows’ protected storage and are kept out of prompts, logs and relay traffic.
Known limit, stated plainly: protection is only as good as each tool’s own classification. On the desktop, keys are classified as keys, not by their effect, so pressing Enter in an app you have allowed counts as typing even if that app then sends something.
06 Privacy
Activity observation is off until you turn it on. Then it notes which program is in front and for how long, not what you do in it.
07 Status
08 FAQ
Not yet. It is at version 0.1.0 and in development: nothing is signed, packaged as an installer or distributed.
Routine steps inside what you’ve granted, such as reading, navigating and typing into an app you allowed, run on their own. Writing files and running programs ask by default. Sending, publishing and deleting ask unless you created an exact standing rule; buying, installing, security and credential changes always ask. Under Suggest, nothing runs except observing and drafting.
No. It makes an outbound connection to the relay and nothing on it listens for inbound connections, so there is no port to open or forward. Each command from a paired device is signed and end-to-end encrypted, and your computer re-checks it against that device’s permissions and its own policy.
Aporion uses Anthropic’s Claude with your own Anthropic API key, kept in Windows’ protected storage. Anthropic bills you for that usage; Aporion’s per-task and daily budgets stop work when a limit is reached.
Not yet. macOS is designed for, but not built or tested.
Yes. The relay is a small service with a container setup and a deployment guide. It routes sealed messages between your devices and your computer, and cannot read or forge them.
Stop is always one step away: a global shortcut, the tray, Stop on every task, stop from a paired device, and an automation-off switch that survives a restart. Anything that was already mid-flight is reported honestly.