Runs where you decide. Learns how you work, and writes it down.
✓ Got it. You’re on the list.
| Opportunity | Stage | Amount | Close |
|---|---|---|---|
| Acme Corp | Negotiation | €48,000 | 30 Sep |
| Nordwind Logistics | Proposal | €22,500 | 14 Oct |
| Bellamy Studio | Qualified | €9,800 | 21 Oct |
| Kestrel Health | Discovery | €61,000 | 2 Nov |
| Orbit Media | Proposal | €15,200 | 9 Nov |
Your assistant can tell you which menu to use. It can’t click it. So you do the work anyway.
The work is done. Now there’s a second job: telling a system it’s done. Forget once and the record is wrong.
Your assistant was in the meeting too. It knows what was decided. Nobody else does until you type it up.
How your team works lives in three people’s heads. When they’re busy, everyone waits. When they leave, it leaves.
The tools that would help send your work to someone else’s server. Your contracts say no.
A hundred emails a day, Slack on top. You skim, and still don’t know what’s going on.
Each app opens in its own tab. Beside it, a chat that can operate that app.
Where an app has an API, the assistant uses it. Where it doesn’t, it works the app itself: the same menus, panels and shortcuts you use. Nothing is out of reach.
A master chat works across every tab at once. Ask once, or leave standing instructions.
A task three menus deep, done through the app itself.
One request, several apps, in parallel.
Every action starts from a fresh look at the screen: this window, this state, right now. An old screenshot can’t authorise a click.
Before input, a separate check confirms the target, the permission, and what’s at stake. After, it confirms the result, and that nothing else moved.
Clicked is not done. It tells you which one happened.
Allowed, delivered, verified. Three separate answers.
If an action may have gone through but the app didn’t confirm, it checks instead of trying again. One email, sent once.
Undo isn’t a safety net. Ctrl+Z can’t unsend a message or bring back an overwritten file, so anything that can’t be taken back waits for your yes.
Grab the mouse and it lets go. It works beside you, not over you.
Uncertain means check, not repeat.
Do something twice and it becomes a procedure: readable English you can open, edit, and run again.
Every action is logged as a named operation, timestamped, attributed to a model.
Procedures are versioned like code. Your method, under version control.
Round off the edges of whatever is selected without changing its silhouette.
A procedure is a readable file, under version control.
Named operations, timestamped, attributed to a model.
One person figures it out. Everyone’s assistant learns it.
Procedures are shared. Someone solves a problem on Monday; on Tuesday everyone has it, on their own machine.
New people inherit the method instead of relearning it.
There’s a terminal and an editor in the box. When a job needs a script or a small tool, the assistant writes it, runs it, and keeps it for next time.
Software shaped like your work, without a ticket or a contractor.
A one-off tool, asked for and built in the same tab.
Hermes, Prime Intellect, OpenCode and others ship with it. Pick one per tab, or swap mid-task.
Models come through OpenRouter, through us, or from your own machine. Nothing ties you to one vendor.
| Tab | Agent | Model | Where |
|---|---|---|---|
| Blender | Hermes ▾ | via OpenRouter ▾ | cloud |
| Salesforce | Prime Intellect ▾ | via Drive ▾ | cloud |
| Terminal | OpenCode ▾ | Local ▾ | this machine |
| Kdenlive | Hermes ▾ | Local ▾ | this machine |
| GIMP | Prime Intellect ▾ | via OpenRouter ▾ | cloud |
Agent and model, chosen per tab.
Local, cloud, or both, set per tab. Shared work goes to the cloud so procedures move between people. Sensitive work stays on the machine it started on.
Local is not a degraded mode. Pull the cable mid-task and it finishes.
Cable out, task finishes.
Apps with a good API are quick. Everything else the assistant operates directly, which is slower. Coverage is the same. Speed isn’t.
You set where each tab runs and where its data goes.
It versions the method, not the assets. Your files stay wherever they already live.
Windows, macOS and Linux.
Only if you pick a cloud model for that tab. Run local models and nothing leaves the machine. You can check by pulling the network.
Where you do use a provider, you set the data retention rule for it. Different providers, different rules, decided by you rather than inherited.
In the settings app, inside your machine. Tokens are encrypted at rest in the operating system’s own credential store and never written to a skill file, a log, or a model prompt.
Assistants request a connection rather than reading the token. You grant it per app, you see what was granted, and you revoke it in one place. Nothing is held on our servers.
Only the ones you grant, one window at a time. A window with the same title isn’t the same window, so it asks again.
Each chat has its own access. One chat can’t see another’s screen or borrow its permissions.
It steps back. Working alongside you is a mode you switch on, so it never fights you for the cursor.
Yes. Anything that installs on a normal machine installs here, and the assistant can operate it once it’s in. Licensed software comes in on your own licence.
They’re plain text files you can read and take with you. Files stay in standard formats.
Per seat, plus storage. Model usage is billed at cost when we’re the provider.
Runs where you decide. Learns how you work, and writes it down.
✓ Got it. You’re on the list.