An autonomous operator for your computer.
Desktop agent, perception, plan, action.
Native AI,built like instruments.
We build local-first instruments that observe before acting, ask before changing, and leave a trail you can replay. On your machine. With your consent. No noise.
A desktop assistant that sees, decides, and operates.
An AI companion that observes the shared screen, understands what's on it, and, only after you allow it, drives the keyboard, the mouse, the windows, the apps.
- Always there
A floating widget, one keystroke away.
Curie lives as a small desktop companion: discreet when idle, ready when summoned. Pin it on a corner of your screen and call it during real work, not in a separate chat window.
- Sees the screen
Multimodal vision with a coordinate grid.
Curie reads a captured frame with a precise coordinate grid overlaid. Targets are picked from the grid, never guessed, so a click lands on the pixel it was supposed to.
- Acts safely
Click, type, drag, only with proof.
Mouse moves, key sequences, app launches, window focus, scroll, drag. Every action returns evidence: foreground window, cursor position, verification capture, clipboard content.
- Routes the right model
Internal LLM router. Vendor-agnostic.
Curie's backend orchestrates between models: a fast model for micro-decisions, a strong vision model to read the screen, a deeper reasoner for hard plans, an economical model for non-critical work. The product evolves with the whole LLM ecosystem.
Press a key. Get the prompt you meant.
A keyboard companion that turns sloppy intent into a structured prompt. Asks two questions, max. Lives one shortcut away from any text field, running on your own machine.
- Anywhere
One shortcut, in any text field.
Press ⌥ Space inside Slack, Notion, Mail, browser, terminal, Laxus rises as a quiet panel rooted to the field you came from.
- Listens for intent
Type sloppy. Laxus parses verb, target, audience.
Laxus identifies what you actually want and stops there. No chat, no scaffolding. It detects what's missing and asks at most two questions to fix it.
- Outputs a brief
Hand back the prompt you meant.
The enhanced prompt drops back into the field you came from. Clear voice, explicit constraints, no fluff. Original kept in history.
- Runs on your laptop
Runs entirely on your machine.
A model small enough to run on your own machine. No request leaves your laptop. No vendor sees what you're writing.
A private AI host that speaks the OpenAI dialect.
Atlas is the private AI host that powers Curie's local path and any client that already speaks the OpenAI API. One install via Toolbox. Free, with zero egress by default.
- OpenAI-compatible
One endpoint. Every app.
Atlas exposes the same chat/completions, embeddings and models routes you already speak to. Point your client at localhost, no SDK swap, no rewrite.
- Models
Pulsar, plus any model you bring.
Atlas ships with Pulsar, the Scalaris vision / agent / text model, and runs your own models through the same manager. Vision, chat and embeddings, all served locally, nothing leaves your machine.
- Local-first
Zero egress by default.
No port is opened on the network unless you say so. No telemetry. No model card phones home. Atlas is the side-car you can audit, not a SaaS.
- One install
Toolbox installs and supervises.
Atlas is shipped via the Scalaris Toolbox. One click, the runtime is on disk, registered as a session service, and visible in your account.
A terminal cockpit, built around real shells.
Termiris turns the terminal into a cockpit: a grid of real OS shells, projects that survive a reload, LAN-only mobile pairing, a live-screen overlay, and widget boards on every spare display. Real operating-system shells everywhere, never a scripted demo shell.
- Real shells
A real TTY in every pane.
Every pane is a real operating-system shell. claude, gemini, aider, ssh, vim, htop run untouched, resize signals, mouse pass-through, alt-screen and true color all carried, never scripted.
- Modular grid
Splits, projects, layouts that survive a reload.
Drag the hairline dividers to split horizontally or vertically, ratio clamped, instances cached so a route change never kills a shell. Each project persists its layout and per-cell spec on disk; switching projects leaves running shells alive in the background.
- Pair · overlay
Mobile remote on LAN. Live-screen on a click-through canvas.
Pair your phone via QR over your local network, with a countdown, nothing leaves your devices. The overlay paints a live screen preview into a transparent, click-through window, so it floats over real work without ever intercepting a click.
- Widget boards
Every spare display becomes a cockpit.
Termiris spawns a borderless fullscreen window on each non-primary display. 12 × 8 grid, snap guides, edit mode one keystroke away. Seventeen widget kinds out of the box: hero clock, GitHub contributions, CPU/RAM/disk, weather, Pomodoro, RSS, crypto, calendar, todo, world clocks…
An applied AI lab.
Scalaris is an applied-AI lab building software that sits next to your real work, on the machine in front of you.
We turn research into binaries you can run today, fewer and more precise things, private by default.
We believe the next decade of agents will be measured in milliseconds on local hardware, not tokens in a datacenter.
We hold the lab to the same bar we hold the agents to: be quiet, be exact, leave a trail.
[ Get in ]
Pick a product. We'll take it from there.
Both products run a closed alpha. Twelve to eighteen people every month. Tell us which instrument you'd reach for first.
EU, encrypted at rest, deletable on request.