The thing that kept me from trusting AI agents — back when I was just a user of them — wasn't that they made mistakes. Everything makes mistakes. It was that I couldn't see the mistakes happen. An agent would churn for ninety seconds and hand me a result, and I had no idea whether it had done the right thing for the right reasons or stumbled into a plausible-looking answer.
Froots is my answer to that problem. The tagline is "multiply yourself," but the principle underneath it is narrower and more stubborn: you should be able to watch an agent work, and stop it before it does something you'd regret.
Visibility is the feature
Froots is a local-first desktop app for running AI agents — Claude, GPT, Gemini, local models, or a custom harness — with full visibility into every step. Tool calls, file edits, browser sessions, and the model's reasoning all stream into a single timeline. Not a spinner. Not a summary after the fact. The actual sequence of actions, as it happens.
This sounds like a UI detail. It isn't. It changes the relationship you have with the agent:
- You catch a wrong turn at step three instead of discovering it in the output at step twelve.
- You learn how the agent actually reasons, which makes you better at directing it.
- You can hand it riskier work, because "risky" stops meaning "blind."
Interruptible by design
Visibility without control is just a nicer way to watch something go wrong. So Froots lets you pause a run before a risky action, review the proposed change, and approve or reject it. Edits to files, commands about to run, a form about to be submitted — these are checkpoints, not surprises.
I think of this the way I used to think about risk in trading. You don't refuse to take positions; you size them and you set stops. An agent should be free to act, but the irreversible actions should pass through a gate where a human can still say no.
Local-first, your keys, your machine
Froots runs on your computer. You bring your own API keys, so you're paying model providers directly rather than paying me a markup to sit in the middle. Your files stay on your disk. Syncing across devices is the one optional paid layer, and it's there because people asked for it — not because the product is hostage to a subscription.
That decision is partly philosophical and partly practical. An always-on agent that touches your real files, your real browser sessions, and your real accounts is not something you want renting space on someone else's server by default.
Multi-panel, because one agent is rarely enough
Real work fans out. While one agent refactors a module, another can be researching in the built-in browser, and a third can be drafting the summary. Froots runs them in separate panels, each with its own timeline, so parallel work stays legible instead of turning into one undifferentiated stream.
Why I'm building this way
I've shipped products where the magic was hidden on purpose, and I've come to believe that for agents it's exactly backwards. The trust you need to delegate real work doesn't come from the model being impressive. It comes from the system being legible — from reliable memory you can inspect, from browser automation that reports what it actually did, and from a timeline that never asks you to take its word for it.
That's the whole bet. If you want to feel it, try Froots.
Dylan Worrall is the founder of Froots and Soshi Labs. He writes about agent architecture and the engineering behind dependable AI products.