An AI teammate inside your browser, not just a chatbot.
One side of the OS: an agent that works on the pages you’re already on — inside your own logged-in browser. It reads, acts, researches, and remembers, all approval-gated and cited.
One agent. Many tools. Every surface.
A model answers questions. An agentic OS gets work done: it understands what you are looking at, decides which tools to use, takes action with your approval, and carries context across the apps and devices you work in.
The agent loop
A single server-side reasoning loop drives every task — choosing tools, reading observations, and streaming results — with step and cost ceilings as circuit breakers.
The tool layer
Page understanding, browser actions, research, memory, workflows, and files are tools the one agent composes — each permission-checked and classified by risk.
The clients
Hands and eyes live in the client you use — the browser extension and web app today; desktop and mobile next — while the agent and your data stay consistent across all of them.
What the one agent can actually do
Each capability is a first-class tool — built to be reliable on its own and powerful in combination.
Understand any page
The agent reads an accessibility-first, distilled view of whatever you are looking at — web pages and PDFs alike — and answers questions grounded in the actual content, citing the exact source lines. It never dumps raw HTML and never guesses.
- ›Accessibility-tree snapshot, not a brittle HTML scrape
- ›Answers cite verbatim source passages
- ›Reads on-page PDFs and documents
Fill & submit forms
Point the agent at a form and it fills fields precisely using stable element references. Anything consequential — submitting, sending, paying — pauses for your explicit approval first. Sensitive fields like passwords are never read or filled silently.
- ›Stable element targeting, realistic input
- ›Consequential actions always confirm first
- ›Refuses sensitive fields by default
Do multi-step tasks
The agent acts on your behalf — navigate, click, type, select, scroll — taking a fresh look at the page after every step so it stays grounded. You stay in control with a live action log and approval gates on anything irreversible.
- ›Re-reads the page after each action
- ›Live action log you can watch
- ›Reversible by design, gated when not
Research with citations
Ask a question that spans the web and the agent opens, reads, and synthesises multiple sources into a numbered, cited report — within a per-task source budget you can extend whenever you want to go deeper.
- ›Numbered [n] citations back to sources
- ›Per-task source budget you control
- ›Public-source fetching is SSRF-guarded
Work across your tabs
The agent can read the tabs you already have open and open new background research tabs, then bring everything together — useful for comparing options, reconciling data, or summarising a session of reading.
- ›Reads tabs you already have open
- ›Opens background research tabs
- ›Cleans up the tabs it opened
Remember what matters
Tell the agent something once and it remembers — your preferences, the people and projects you work with — recalled automatically when relevant. Memory is strictly scoped to you, refuses secrets, and is yours to search or forget at any time.
- ›Semantic recall, scoped strictly to you
- ›Refuses passwords, cards, and secrets
- ›Search and forget anything, anytime
Routines — teach it once, re-run forever
Do a task once with the agent watching, then save it as a routine. It captures the steps, auto-names them, and turns the values you typed into fill-in-the-blank inputs. Replay it later with new inputs — across hundreds of rows or on a schedule. When the site changes, it self-heals, re-finding the right element on its own, and pauses for you when it genuinely can’t — never silently skipping ahead.
- ›Record once — steps and inputs captured automatically
- ›Replay with new inputs, at scale or on a schedule
- ›Self-heals when a site changes; pauses when it can’t
Work with your files
Attach a document and the agent reads it, pulls out what you need, and can edit text and Word files — the edited version downloads as a new file. Your original is never overwritten.
- ›Read & extract from PDFs and documents
- ›Edit text and Word files in place
- ›Originals are never overwritten
Fix writing in place
Proofread any block of text and get a corrected version with the edits highlighted in place — spelling, grammar, and punctuation fixed, one click to copy. Re-run to refine.
- ›Corrections highlighted in place
- ›One-click copy of the result
- ›Re-run to refine the wording
Teach it once. Re-run forever.
The bridge from a one-off task to real automation. Do a task once with the agent watching, save it as a routine, and it becomes a reusable, fill-in-the-blank recipe you can replay — one run, hundreds of rows, or on a schedule.
Record once. Do the task with the agent watching. It captures every step and auto-names the routine.
Fill-in-the-blank inputs. The values you typed become parameters — replay with new inputs each time.
Self-heals when a site changes. It re-finds the right element on its own when a page’s layout shifts.
Pauses when it can’t. It never silently skips ahead — it stops and asks, and gates anything consequential.
Built on principles, not promises
The things that make an agent safe to hand real work are designed into the core.
Consent-first
Reading and reversible steps flow freely; consequential actions pause for explicit approval. You are always the one who decides.
Grounded & cited
Every answer about a page or the web is built from real content with verbatim citations — never a guess, never a hallucinated source.
Private memory
Memory is semantic, strictly scoped to you, refuses secrets, and is fully searchable and erasable. Yours, and only yours.
Scale-safe
Runs persist and resume from durable state, so a task survives closing the panel and the system scales horizontally without losing your place.
Put the browser agent to work.
Start free on the web and run your first routine in minutes. No card required.