An autonomous operator for your computer.
Desktop agent, perception, plan, action.
Curie sees what's on your screen, plans the next move, and operates your computer through pixels and inputs, locally on your machine.
Four steps. 142 ms median. Every cycle is hash-stamped, on-device, replayable.
Twelve frames per second of structured perception. Pixels become tokens, on-device. No extension, no accessibility tree, no API contract.
12 fps, 0 B egress, on-deviceA local model decomposes the goal into a small executable graph. Every step is checkpointed. Every plan is replayable.
on-device, 6 nodes, checkpointedPointer, keyboard, voice. Curie operates your apps through the same surface a human uses. No API required. Scoped per-app.
pointer, keyboard, scopedHash-stamped. Re-run a session, diff between runs, share with a teammate, export for compliance. Time-travel is built in.
hash-stamped, diff, exportA small stack of replaceable models with a thin orchestrator. Anything that touches the network is opt-in, scoped, and logged.
Combined into anything you can describe. Six primitives, each one documented, each one tested across hundreds of apps.
Mail, Notion, Figma, Slack, Excel, Linear, Calendar, Terminal. Curie operates them through the same surface a human uses: pixels and inputs.
Forms, scraping, dashboards, multi-tab flows. Curie reads pages like a person. No extension, no API, no fragile selector.
Long-running shell tasks with safe-mode constraints. Drives git, builds, deploys. Reads stack traces, fixes them, opens the PR.
Every action gets a deterministic trace. Re-run a task, diff between runs, share a session, audit a decision. Time-travel for the agent.
Long-lived workspaces. Curie remembers your tone, files, contacts, your taste, only for the projects you choose. Ephemeral by default.
Talk, point, gesture. Curie listens and watches at the same time, with glanceable feedback. Voice stays entirely on your machine.
End-to-end, from screenshot capture to confirmed input. Measured across 1.2 million actions on a passively-cooled M-class machine.
Internal, 2026 Q1, n = 1.2MRe-run any session and get the same actions, bytes, and timings. Action-graph hashes are 100%; pixel-level state is 99.94% across hardware.
10 000 sessions, self-replay diffCurie ships with networking off for inference. Audit-only mode is opt-in and signs every uploaded byte. The cloud is a feature, not a dependency.
tcpdump, idle-to-actionTwelve to eighteen people every month. Tell us what's open on your desk right now.