Browser Extension · Web Surface

The autonomous browser agent.

A Chrome Manifest v3 extension that drives the browser by voice. Reads the active tab DOM, listens for prompts, plans browser actions, executes click, type, scroll, and form fill, and speaks Model Context Protocol to local MCP servers on your machine. Ask it from your phone to find the cheapest flight next Friday and walk away while your desktop does the work.

INSTALL EXTENSION API DOCS

Chrome Manifest v3

Real MV3. No legacy hacks.

Built ground up for Chrome Manifest v3, with service worker scripts for orchestration, content scripts for DOM access, and declarative network rules for clean request handling. No abuse of long lived background pages. No legacy permissions wedge. It installs cleanly on Chrome, Edge, Brave, and every Chromium fork without warnings or compatibility shims.

MV3 service workers, content scripts, and offscreen documents
Installs on Chrome, Edge, Brave, Arc, Opera, and Chromium
Minimal permission surface, scoped to active tab and audio capture
Auto update through the Chrome Web Store

ARCHITECTURE DOCS

Chrome MV3 · service worker

⚙manifest v3

DOM · read

☰active tab context

DOM context

Reads the page. Understands the page.

The extension extracts a structured view of the active tab DOM, identifies interactive regions, headings, forms, buttons, prices, dates, addresses, and feeds it into the Orb model layer for planning. The model does not see a flat blob of HTML, it sees an agent ready map of what is on screen and what the user can do with it.

Structured DOM extraction with semantic region tagging
Form schema inference, button label normalisation, price detection
Visual aware element ranking for ambiguous queries
Same NRN memory layer as every other surface

DOM CONTEXT DOCS

Plan and execute

Click. Type. Scroll. Submit.

Voice in, plan out, actions executed. The extension turns prompts into ordered browser actions: click a button, type into a field, scroll a list, fill an entire form, navigate a flow, capture a screenshot, return a result. Every step is observable. Every step is revocable from the Orb on any other surface. Long horizon flows can run while you are on a walk.

Action graph: click, type, scroll, hover, drag, form fill, screenshot, wait
Per step observability: every action is logged with a screenshot
User confirmation gates for destructive or financial actions
Pause, resume, cancel from the phone, watch, terminal, or wallpaper

ACTION DOCS

action graph

➤plan and execute

MCP · local

⊖model context protocol

Model Context Protocol

Speaks MCP to your local machine.

The extension is an MCP host. It speaks Model Context Protocol to local MCP servers running on the user machine, bridging the browser to your filesystem, your shell, your local APIs, your database tools, and any other MCP server you have running. Browser sessions get full local tool access, with explicit allow lists and per server scopes.

Full MCP host implementation: tools, resources, prompts, sampling
Connects to local MCP servers over stdio or local socket
Per server allow list, per tool scope, audit log of every call
Bridges browser tasks to filesystem, shell, and local tooling

MCP DOCS

Drive from your phone

Speak from your phone. Browser does the work.

The killer demo. You are on a walk. You ask the Orb on your phone to find the cheapest flight to Cape Town next Friday and a hotel near the V and A Waterfront under R2500 a night. The desktop extension takes the prompt, opens tabs, fills searches, compares results, drops the top three into a shared note, and pings the watch when it is done. You arrive home, the work is finished.

Voice prompts from any Agentics surface dispatched to the extension
Live progress, screenshots, and intermediate results streamed back
Cross device handoff: start on phone, observe on watch, resume on terminal
Sandboxed scopes per task, automatically expire on completion

REMOTE DISPATCH DOCS

phone → browser

✈autonomous shopping

Stop browsing. Start dispatching.

A Chrome MV3 extension, an MCP host, and an autonomous agent in one. Speak to your phone, walk away, find the work finished when you get home.

INSTALL EXTENSION READ THE DOCS