Skip to content
Prototyper memo
date:
thu, jun 18, 2026
to:
everyone
from:
founding team
subject:
our company

Prototyper's mission is to make every company agent-native.

A new kind of company is forming. Agents now do work that used to require a teammate — they write code, draft specs, search the company's data, ship pull requests. They don't live in a sidebar; they live in the work. The companies that will pull ahead in the next decade are the ones that treat agents as first-class workers and design their tools around that fact. That's what we mean by agent-native.

The tools to make that real don't exist yet. The cost of finding out which software ideas are worth building is still too high — designers draw pictures, engineers rebuild them, six weeks later the answer comes back, and the answer is almost always that you should have killed it. So you ship anyway, because the alternative is admitting the time was wasted. Most ideas are wrong. For every yes there are a thousand no's. We've watched this loop bleed companies of their best work for a decade.

The technology to end the loop finally exists, and the architecture writes itself once you see what it's built on. Every coding agent shipping in production — Claude Code, Codex, Cursor — converges on the same primitive set: read, write, grep, glob, bash. They reason in unix because they were trained on it. Their natural substrate is the filesystem; discovery happens via grep and ls, not via a schema we stuff into the system prompt. The agents already speak this language. We just had to build the room.

So Prototyper is, plainly: agents as processes, the filesystem as inter-process communication, the canvas as the display. A multi-agent operating system with a spatial display layer. Single-player versions of this idea have shipped — Claude Code, Cursor, the Codex CLI. Nobody had made it multiplayer. So we did.

We designed Prototyper from first principles:

  • Agents are first-class workers. Not assistants in a sidebar. They draft specs, write code, search the company's data, open PRs. Same surface as the humans, same files, same accountability.

  • One canvas, many rooms. Design today; agent orchestration, knowledge work, operations tomorrow. The substrate doesn't care what you put on top of it. Different rooms, one company.

  • Bring your own AI. Plug in the Claude Code, Codex, or Cursor subscription you already pay for. No margin on top of someone else's tokens. Your model, your pricing, your choice.

  • Every artifact is a file. Skills, plans, memory, conversations, agents, mounts — every artifact inherits the same operations: sync, mount, version, replay. The plumbing collapses into one shape.

We're starting with the surface we know best: the canvas where designers, engineers, and agents ship the actual product, not pictures of it. But the substrate doesn't care what room you put on top of it. The same kernel runs agent orchestration, knowledge work, analytics, anywhere else a company wants its people and its agents on one surface. Different rooms, same substrate. We'll ship the rest one room at a time.

I've spent the last decade between design and engineering at startups and platforms, watching how much time gets lost in the gap between what gets drawn and what gets shipped. That gap is the most visible symptom of companies that aren't agent-native yet. The faster you can rule out a no, the sooner you find the yes — and the sooner the company finds its real shape.

First proof: more than 10,000 real React apps have shipped from the canvas — designers, engineers, and agents working the same surface, on the same files, in production. From the same substrate we maintain three open-source projects — tmux-ide, prototyper-ui, vibecode — public, MIT-licensed, free to fork. If we're going to ask the community to build with us, we have to be the ones building first.

We hope you'll trust us with your work.