Ask it anything — in your browser, in any desktop app, on your phone — and a guide-cursor glides to the right element and shows you what to do, step by step. The AI never clicks. The human always acts. It's GPS for software: never scolds, always reroutes, and steps back as you learn.
A parent stuck on a railway-booking form. A new employee lost in the company's CRM. Anyone facing an unfamiliar app with something at stake. Their options today all fail the same way — none of them teach.
The user performs every action with their own mouse.
Success for Second Cursor is the moment a person no longer needs it for that task. This is also its safety guarantee — an assistant that cannot click cannot be tricked into clicking, cannot buy anything, cannot delete anything. It can only point, explain, and encourage. That makes it deployable where autonomous agents can't go: elders' devices, regulated workplaces, employee onboarding, education.
One hotkey (or the caret button) summons the command bar. Below, five asks on five surfaces, played one after another: a train booking in the browser, a file created from the Windows command line, a setting buried deep in macOS, a recorded IT workflow replayed on Linux, and a video call on Android. The same brain guides all of them. Pick a surface, or click any step to jump.
Interruptions are first-class: popups, cookie banners, and surprise dialogs are classified on the fly — absorbed into the plan if task-relevant, calmly dismissed if blocking noise, ignored if harmless. Tab or app switches pause the session; returning resumes it exactly where it was. Max three reroutes, then a graceful, friendly stop — never an error message, never blame.
Second Cursor tracks how well you know each task — how many times you've been guided, how fast, how many mistakes — privately, on your own profile. That mastery score drives a deliberate teaching progression:
Halo on the exact element, plain-language instruction, every action verified. Every tooltip carries a tiny ? — tap it and the step explains why, so understanding transfers to other apps.
Same task asked again? The guide notices and steps down: halo only, shortened instruction. It stops spelling out what you already know — and tells you so, warmly.
Roles flip: you do it, it watches. The cursor parks; a silent watcher checks each action against the plan and steps in only on a mistake or a 10-second hesitation. Finishing solo is celebrated — that task is yours now.
Not “do this for me” but “what am I looking at?” Ask “is this email a scam?” and the cursor glides between the actual red flags on your screen, narrating each in plain words, advancing when you tap “next”. Same engine, zero actions.
Hesitation patterns — mouse circling, hovering without clicking — are noticed on-device, and the guide gently offers help before being asked: “want me to show you again, slower?”
Colorblind-safe high-contrast skin, cues never color-only, an elderly preset (2× size, slower pace, dock-bar mode, 16px click tolerance for tremor), screen-reader coexistence — the overlay never steals focus.
The first person guided through a task pays the planning cost once. The validated step-list is saved as a recipe — every later ask for the same task replays instantly, with zero model calls.
Recipes are self-healing: they match screens by meaning, not pixel positions, so redesigns still work — and when an app changes enough to break one, the system replans once and the recipe updates itself. Each variant carries a trust score that grows with confirmations and decays with age.
This is the moat: a growing knowledge base of how software is actually operated — faster and cheaper with every user. Proven in the end-to-end suite: the second identical ask completes with 0 provider calls.
The identity is a matrix-green terminal caret — deliberately unlike any native cursor, so guidance is always unambiguous. That default stays free forever, alongside the colorblind-safe high-contrast skin and the elderly preset (2× size, slower pace).
Paid skins go well beyond colour: animated companions that dance beside the caret, character packs, seasonal and limited drops, team-branded cursors for organizations, and creator skins with a revenue share. Cosmetic-only by design — a skin never changes what the guide can do, so the safety story is untouched.
To guide, we must see the screen's structure — where the buttons and fields are and what they're called. Here is exactly how that never becomes surveillance:
The snapshot's value field is hard-coded to empty for every input, always. Typing steps are verified by the transition “empty → filled” — never by content.
Stored: “the Compose button leads to a form with an attach control.” Never stored: who did it, what they wrote, any content beyond element labels. Shared knowledge passes through an anonymizer.
Password and payment fields are known only by role. On payment or final-confirmation steps the guide adds a caution note and never auto-advances.
Struggle detection runs on-device; only the derived signal (“stuck on step 2”) ever leaves. Feedback is anonymous and opt-in. Mastery progress is private to your profile.
Screen content enters the model as explicitly untrusted data; plans are schema-locked to on-screen elements. Safety rules around payment live in static code the model cannot override.
Model credentials live in the brain's local config — never in the extension, never in a web page's reach. In managed mode, keys live server-side only.
Read any screen into a universal, privacy-filtered map. Draw the cursor, halo, and tooltips. Watch the user act and verify each step on-device in under 100 ms.
It can see and show — but holds no intelligence, no data, no ability to act. Swap browser for Android or Windows and nothing else changes.
Understands the question, chooses the cheapest planning path, verifies outcomes, manages the session — and owns the recipe knowledge base that makes repeat tasks free.
This is the company: the accumulated know-how, the trust system, the economics. It never knows which platform it's serving.
Claude, GPT, Gemini or an on-device model — called only when a task is genuinely new, through one narrow, injection-armored interface with strict structured output. Users can bring their own key from a curated list.
Models are a supplier, not a dependency — as recipes accumulate, model spend per task trends toward zero.
The endgame is the same guide living OS-wide: a quiet presence beside your real cursor across every application, backed by the same knowledge — with cross-device continuity: start a task on the phone, finish it on the PC. Sessions are device-agnostic objects in the brain, so any adapter can render the next step.
Three ways to run it, one product, plus skins on top. The unit economics improve with usage: recipes make repeat tasks free, so serving a user gets cheaper the longer they stay.
We hold the model keys; the user just installs and asks. Free tier for light use, subscription for unlimited guidance. Helper setup lets a relative install remotely; a caregiver view (later) shows a linked elder's progress.
Your own API key from a curated model list — you pay your provider, we charge little or nothing for the software. This tier seeds the community: BYOK users are the natural first workflow authors and recipe contributors.
Companies buy living onboarding: record workflows once, pin them into a team pack, every new joiner gets guided hands-on training inside the real tools. The same machinery deflects support tickets.
Matrix green is free. Premium and exclusive skins, limited drops, team-branded cursors and creator skins with revenue share. Cosmetic only — never changes what the guide can do.
The core loop is built and running — real questions, real model planning, real guidance on live websites. Labels are literal: BUILT = running and tested today · DESIGNED = fully specified, next in build order · PLANNED = roadmap.
automated tests green across brain, extension, protocol and end-to-end suites — including live-site regression and a 5-pattern showcase gate
model calls on a repeat ask — the recipe learning loop is proven in the end-to-end suite
per-step verification, fully on-device — guidance stays instant even with a cloud brain
functions to support a new platform — see the screen, draw the cursor, watch the user
Chrome extension + local brain. Any-website guidance, the recipe loop, cross-page tracking, the full tutor layer, settings with live pickers, accessibility skin — closed by a 5-pattern end-to-end showcase gate.
The brain deploys unchanged to the cloud; testers install only the extension. A shared recipe database becomes the first live proof of knowledge compounding across users.
Sign-in and cross-device sync, recorded workflows, packs & team sharing with quarantine trust, voice in/out, the Android adapter, guardian protection as standard.
Windows/macOS/Linux adapters — the guide OS-wide beside your real cursor — iOS, cross-app workflows, a public pack library, skin store, elderly onboarding as a product of its own.
Every guided task makes the next user's identical ask instant and free. Cost per task trends toward zero — the opposite of per-task agent economics, and a data asset competitors can't shortcut.
The agent race replaces the human at the controls. Second Cursor serves everyone who wants — or needs — to stay at them: elders, regulated workplaces, onboarding, education. Can't-click is a safety story automation cannot match.
Intelligence, knowledge base and trust system carry unchanged from browser to phone to desktop. Each new platform is a thin adapter measured in weeks.
Recorded workflows are user-generated content for software know-how. Families, communities and companies fill the library — we verify and distribute.