About

A small team building a desktop AI company.

We think the wedge in agentic AI is not capability — it is hiding the wiring and showing the work. This page explains the bets we are making and the principles we keep coming back to.

Why a desktop app?

Most AI products today are either single-agent chat (great UX, no real coordination) or developer agent frameworks (real coordination, no UX). We think there is a third shape: a desktop app where a non-technical user runs a hierarchy of AI employees the way they would manage a small team — by DMing people and letting managers route work.

Desktop is the right surface because the work is local: the App Builder writes files into a folder the user can see, the filesystem and shell are real, and nothing about a chat thread with an employee needs to be cloud-synced to be useful.

The validation question

“Did running an AI company feel like managing a small team without learning agent infrastructure?”

Not “is any one employee useful?” — that bar is met by every other chat app. The bar we care about is whether the coordination model feels natural to a non-technical user the tenth time they open the app.

Principles

Six bets we keep coming back to.

The agents are the product
The chrome should never compete with what the agents are doing or saying. Monochrome shell, agent identity carries the color.
Hire, don't configure
Users pick employees from a catalog, not models, prompts, or tool lists. The org tree is built from each template's reports_to_roles.
Safety budgeted per task
Each task has token, wall-clock, and USD caps. Delegation is depth-capped. Destructive actions require inline approval.
Coordination over capability
Multi-agent reliability is a known industry problem. We contain it via depth caps, parallel-delegation caps, and explicit cancellation.
Hide the wiring
API keys, MCP servers, tool names, JSON schemas, token counts — all hidden by default. The user thinks in jobs-to-be-done, not agents-to-configure.
Local-first
All state lives on the user's machine. No prompt or response content leaves the device in dev or beta. Future telemetry ships only structural events.
Private beta