Sai

I'm Ang and I build Sai at Simular. Sai is the world's first robosecretary that comes together with a fleet of autonomous computers doing your endless screen work. We believe the future of work isn't just one computer — it's many computer. Each computer navigates software the way you do: reads the screen, clicks the button, types in the form. That's the whole point. Because Sai works at the interface layer, it automates things that have no API and never will — legacy desktop apps, internal portals, anything behind a login. Windows, macOS, and Linux. 73% on OSWorld. What that buys you in practice: 🖥️ Real computers, not a chat box. You hand over a task, it runs it on an actual desktop. Open the workspace any time to watch, or take the mouse back mid-run. 🔁 It learns the route once. Any finished task can be saved as a skill, scheduled, or triggered by an event — so the tenth run is faster and cheaper than the first. 🔒 Autonomy with a seatbelt. Tiered approvals, "trust for this task," and encrypted input for passwords and codes — the model never sees plaintext. The agent we actually optimize for is whether it's still useful and affordable the tenth time you ask. Tell us the screen work that eats your week — we'll tell you honestly whether Sai can take it off your hands.

I'm Chenchen, engineering lead for Sai at Simular. My job is making sure it keeps working after the demo ends. Most AI agents you've seen run on one machine, in one happy path, in a video. Sai runs on a fleet of real cloud computers. Hand it five tasks and five machines wake up, open the apps, read the screens, click, type, and check their own results. That's the product. It's also the hard part. What we actually spent the last six months on: Reliability at fleet scale. Machines sleep, restart, update. An agent that only works when everything is healthy isn't useful. Sai has to wake machines on demand, recover mid-task, and never let two agents grab the same computer. Seeing the screen efficiently. Sai uses Smart Snapshot: it reads what's on screen with far fewer tokens than raw screenshots, without losing accuracy. That's where most of the cost savings come from. Making it cheap enough to give away. A screen-driving agent is expensive by default. We got the cost down far enough to ship a free plan. The number I care about: on OSWorld 2.0, Sai scores 73%, ahead of Claude Opus 5 (70.6%) and GPT-5.6 Sol (62.6%), at $15.70 per task vs their $23.70 and $26.62. Public benchmark, and the cost is what made the free tier possible. Sai has been in production with real users for six months before GA. Ask me anything about running agents at scale. I'll be here all day.

I'm Li Hau, engineer at Simular. I spent years building apps that assumed a human would be sitting there doing the clicking. Then I watched how people actually work: twelve tabs, three apps, copying a value from one into another, all day. The interface was never the product — the tasks were. So we built Sai, a robosecretary, and give Sai a fleet of autonomous computers. Each computer runs command, navigates app, check its own work. What you can do on a computer, can be done by Sai's computer. The hardest and coolest part wasn't one agent. It was many. Running a fleet of autonomous computers in parallel — keeping them coordinated, recoverable, and honest about what they actually did — is where most of our engineering went. I'll be in the comments all day — especially happy to talk about the failure cases.

I'm JC, representing the research team behind Sai. We're releasing Sai for free --- to free everyone from the repetitive busywork of clicking and typing on a computer, so we can all spend time on more fulfilling, creative, and impactful work than operating a computer. "Did you say free? Doesn't Sai use LLMs??" Yes! Sai operates a fleet of autonomous computers for complex personal and business workflows. You delegate a task to Sai, and it orchestrates the best and most efficient models: reasoning models for pulling multiple documents from your Drive, state-of-the-art vision models to click on any pixel on the screen, and efficient verifiers to make sure each computer action is safe and correct for the task. "Then how is free possible?!?" You know what's better than using LLMs? Not using LLMs (at least not for everything). After Sai completes your task once, it learns to repeat the task with 10-100x cost savings in future runs, so we're able to release this free tier for a mass audience along with our paid tier for power users. "Are the machines free also?" Yes, Sai's free tier comes with one remote machine, dedicated just for you for 30min per day. We learned from our invite-only release of Sai since March 2026 that the majority of routine workflows can be completed within 30min per day of machine time. So we've optimized Sai's infrastructure so that you get a full personal machine just like your own. "So how do I get started?" Go to https://www.sai.work/, and Sai, your first robosecretary, will show you how autonomous computers live and work in a new digital world --- giving you more freedom in yours.