Docs · v0.2.0

Olwen can use your computer.

The first Olwen could read your day. This one can act on it.

Until now, Olwen watched — your inbox, your calendar, your tasks — and opened the right surface when you asked. With v0.2.0 he gains hands: with your explicit permission, Olwen can drive your Mac the way you would. Open an app. Find the right window. Send a message. And, crucially, check that it actually worked before telling you it did.

This page explains what that means, how to turn it on, and how to use it.

Two ways Olwen acts

Olwen controls your Mac through a small local helper called the bridge — a daemon that lives only on 127.0.0.1, never the network, and holds a single bearer token on disk. On top of that, Olwen acts in two ways:

1 · Precise, by name

For most tasks Olwen uses the macOS Accessibility API — the same system screen readers use. Instead of guessing where to click, Olwen reads the actual interface and targets elements by their name: "the chat labelled Mom", "the message box", "the Send button". It then types and clicks with the real keyboard and mouse. This is deterministic, instant, and costs nothing — no AI model is involved in the doing.

2 · "Give Olwen the wheel"

For open-ended tasks on apps that don't expose a clean interface, you can hand Olwen the screen. He'll look at it with Claude's vision, move step by step, narrate what he's doing — and then return control to you and report back. A red STOP is on screen the whole time.

Turning it on

Three one-time steps on your Mac:

• 1 — Install  Run the Olwen app (it carries the bridge inside it), or install the standalone bridge from the repo.
• 2 — Grant Accessibility  System Settings → Privacy & Security → Accessibility → turn Olwen on. macOS requires this by hand — it cannot be automated, and that's the point: it's your machine, your consent.
• 3 — Connect a brain  A Claude key in Settings powers "give Olwen the wheel". The precise, by-name control works without any key at all.

What you can say

Just talk to Olwen the way you'd ask a person:

For messaging, Olwen opens the right conversation, types, sends — and then reads the chat back to confirm the message landed. If it can't confirm, it tells you the truth instead of pretending.

Honesty is the feature

The hard part of an agent that touches the real world isn't doing things — it's knowing whether they worked. Olwen never claims success on faith. Every real action is verified against the screen or the interface before he says it's done. "Sent — confirmed it's in the chat" only appears when it genuinely is.

"An action isn't done when you fire it. It's done when you can see that it worked."

Where your data goes

Nowhere it shouldn't. The bridge is loopback-only; it never accepts a connection from outside your machine. Your AI key stays encrypted in your local database. When Olwen takes the wheel, screenshots go only to the Claude key you connected, for that task. There is no Olwen server in the middle.

Limits, honestly

• macOS only, for now. The Accessibility approach is Mac-specific; Linux and Windows are not built yet.
• "Give Olwen the wheel" needs a Claude key with credit, and is paced to respect your API rate limits.
• Some apps expose a thin accessibility tree; for those, the wheel is the fallback.

New in v0.3.0 → Dev Studio: build software while you're away — give Olwen a goal, walk away, get a pull request.

Also new → Chat with Olwen on Telegram — run your day, and start coding jobs, from your phone.

Olwen is source-available on GitHub under PolyForm Noncommercial 1.0.0. Read the full changelog, or learn what Olwen is.

These docs were drafted by Claude, Olwen's assistant · Hand-built in Riyadh · 2026