Thijs here! I'm the founder of Prototyper and today we're launching the first visual canvas built for agents.

Couple of interesting lessons from building the product that I think are worth sharing:

For the agents, everything is a file. In Prototyper, everything from plans, to apps, and diagrams, can be read as a file.

We found that a filesystem is the most natural way for an agent to navigate: it discovers new content and functionality just by traversing the tree.

We kept this layer deliberately thin and unopinionated, because it's the substrate, not the experience. It's the foundation that makes everything work on top.

It essentially operates as a distributed, web-based unix kernel.

Because of that it's built for the agents you already use. People run all kinds of models and harnesses, often several at once.

With this architecture, we don't try to lock you into one, it's built to work with any agent that you like. That means that any agent can read from and write to your workspace.

It's fast. File writes land in under a millisecond. That's not a bogus metric, it's what makes the whole thing feel seamless and like a real extension of your thinking. Thanks to our custom unix kernel.

Visual first. Every file in your workspace can be opened on the canvas. Whether you're an engineer working on a new frontend or a PM building a product roadmap.

It's a real visual workspace for the actual work, not a description of it. As said, it's not a blank canvas with a box of primitives: the canvas represents a real unix system, which is the kind of purpose-built, opinionated experience that makes a product.

In essence, the substrate is generic so the things you build on it don't have to be.

What's most interesting is that by this architecture we found that we can decouple system prompt length from agent capability.

I'm happy to get feedback from the community and see what you all think of it :).