PetBuddy Anywhere — the Chrome extension that puts an AI agent on every website you visit. Coming to the Chrome Web Store.🎙 ~38 tools. Voice, page control, Set-of-Marks hands. Gemini Live speech-to-speech.💬 Talk to websites. One install, rides along everywhere. Built & verified — pending store review.
An AI agent on every website you visit. One install. Talks to you out loud. Does things on the page.
Install the Chrome extension once. An AI voice agent — the astronaut, Hermes — rides along on every website you open. It answers questions about the page, fills forms, navigates, extracts data, and talks to you out loud. Over Gemini Live, Google's speech-to-speech model. ~38 tools. No typing. No chatbot. Just talk.
Or don't take my word for it — Vera is on this page right now. · Tap her, bottom-right, and ask her about the extension.
✓ Gemini Live speech-to-speech • ✓ Works on every website • ✓ ~38 tools • ✓ Set-of-Marks page hands
● The extension — built & verified
🧑🚀
HERMES
Astronaut · Set-of-Marks · 38 tools
1
2
3
4
This is the product. Not a video of it.Hermes rides on every website you visit. The numbered markers are Set-of-Marks page hands — say the number, he clicks. 38 tools. All voice. All real.
What you get
An AI co-pilot that rides along on every website.
Not a chatbot. Not a sidebar. An agent that sees the page, understands it, and operates it — by voice.
🗣️
Voice on Every Site
Talk to any website out loud. The agent answers in real time over Gemini Live — Google's speech-to-speech model. No typing, no chatbot box.
👆
Page Control by Voice
Say "map mode" and every button, field, and link gets a numbered marker. Say the number — the agent clicks, types, scrolls, submits. ~38 tools.
🌐
Works Everywhere
Google, Amazon, your CRM, your bank, your own website — anywhere Chrome goes, the agent goes. It's not limited to one site.
🔒
Your Data Stays Yours
Voice is processed through a single-use ephemeral token. No raw API key on your machine. You control when it listens. Turn it off with one click.
The problem
Websites can't talk back.
Right now, every website is a one-way street.
You read. The website sits there. It doesn't know what you're looking for. It can't answer a question, fill a form for you, or take you where you need to go.
You type. Every interaction — search, forms, checkout, navigation — is you doing the work. Click, type, scroll, repeat.
Chatbots don't fix this. A text box in the corner isn't an agent. It can't see the page, can't click anything, and makes you type instead of talk.
Voice assistants live in one tab. Siri and Google Assistant can open a website. They can't operate it once it's open.
38
tools in the extension
Click, type, scroll, navigate, submit, fill, extract data, take screenshots, open tabs, switch tabs, record macros, play macros, mine data from tables, capture pages, save data — all by voice. The agent doesn't just "chat about" the page. It operates it.
Features
What Hermes can do on any page.
Not a chatbot. Not a sidebar widget. A real agent that sees the page and operates it — the same way you would, by voice.
1
Set-of-Marks Page Hands
Say "map mode." Every button, field, link and dropdown gets a numbered marker overlaid on the page. Say the number — the agent clicks, types, or submits. The marks are durable: a number stays on its element for its lifetime, so scrolling doesn't renumber what you're looking at.
2
Voice-First, Real-Time
Gemini Live speech-to-speech. You talk, it talks back — in real time, with a half-duplex gate that keeps the conversation natural. No waiting for "thinking..." No text-to-speech lag. The mouth moves while the words arrive.
3
Multi-Tab Operator
Open a background tab, navigate, switch tabs, close tabs — all by voice. The agent can work in one tab while you stay on another. "Open the pricing page in a new tab and tell me what the Pro tier costs."
4
Data Miner
Extract structured data from any page — tables, card grids, lists. "Pull every product name and price on this page into a CSV." Downloads straight to your machine. No third-party scraper, no sign-up.
5
Macro Recorder
Record a routine once, replay it on command. "Go to that supplier portal, log in, open today's orders, and read me the new ones." The macro stores durable locators — not volatile mark numbers — so it survives page updates.
6
Confirm Gate
Destructive actions — submit, buy, delete — require a real human "yes." The gate is deterministic, lives in the service worker (not the AI's prompt), binds to a label hash (not the fragile mark number), and survives extension eviction. "Buy now" must not fire on "yeah, I bought one last week."
7
Self-Diagnosis
The extension audits itself. Mic health, speaker health, socket state — all measured from real WebAudio frames and Gemini PCM chunks, not metadata checks that lie green. "Audit yourself" and it tells you what's wrong and attempts to heal it.
8
License-Gated Security
Voice is license-gated through the PetBuddy broker. No raw Google API key ever touches a customer machine. The broker mints a single-use ephemeral Gemini token per session. Revoke a license, the agent goes silent.
The numbers
Built. Verified. Ready.
PetBuddy Anywhere is at v0.18.11. It is not a prototype. Every number below is a real measurement from the live codebase.
390
Verified Checks
Offline guards + live broker contract + real-browser E2E. All green on every push.
21/21
Golden Tasks
Real Gemini Live, real license, real pages — 100% on v0.18.11. First clean sweep.
~38
Tools
Click, type, scroll, navigate, submit, fill, extract, screenshot, macro, mine, save — all by voice.
~17K
Lines of Code
MV3 architecture. Four planes: content, service worker, offscreen, broker. One voice transport.
v0.18.11
Current Version
30+ test suites. Zero orphaned tests. CI on every push + 6-hourly cron for server-side drift.
Install
Three steps. Under a minute.
You don't need a developer. You don't need an API key. You don't need a backend. Install the extension, enter your license key, and talk.
1
Install from the Chrome Web Store
One click. The extension installs like any other — no command line, no config files, no build step. Hermes appears as an astronaut icon in your toolbar.
2
Paste your license key
Click the toolbar icon and paste the licence key from your email — it looks like agent_anw… — then hit Save. The broker validates it once and mints a single-use ephemeral Gemini token. No raw API key.
3
Talk to any website
Open any page. Hermes walks out. Say "hey" and start talking. Ask questions about the page. Say "map mode" to see numbered markers. Say the number and he clicks.
Pricing
One price. Every website.
Pet Buddy Anywhere is its own product at its own price. It is not an add-on to a Concierge plan — you do not need a website, an agent, or an account with us to use it.
Your licence key — it looks like agent_anw… — and the install link arrive by email right after checkout. The extension itself is pending Chrome Web Store review.
Questions
You asked. We answer.
It is a Chrome extension. Install it once, and an AI voice agent — Hermes, the astronaut — rides along on every website you visit. You talk to it out loud through your computer's microphone, and it talks back through your speakers in real time, over Gemini Live. It sees the page, understands it, and can operate it: click buttons, fill forms, navigate, extract data, take screenshots, record macros, and more. Think of it as a co-pilot that lives in your browser — not limited to one site, not a chatbot, not a sidebar widget.
Every website. Google, Amazon, your bank, your CRM, your own business site — anywhere Chrome goes, Hermes goes. The extension injects on all_urls and uses the page's own DOM — it operates the page the same way you do. There is no site-specific integration, no API to connect, no "supported platform" list. If Chrome can render it, Hermes can operate it.
When you say "map mode" or ask the agent to do something on the page, the extension overlays numbered circular markers on every interactive element — buttons, input fields, links, dropdowns, checkboxes. Each marker has a durable number that stays on its element. You say "click 3" or "type hello in 5," and the agent does it. The markers are rendered as CSS-positioned overlays, not DOM mutations — they don't break the page's layout. The numbering is durable: a number stays assigned to its element for its lifetime, so scrolling doesn't renumber what you're looking at.
Yes — you speak to it through your computer's microphone, and it speaks back through your speakers. The audio flows through a WebSocket to Gemini Live, Google's speech-to-speech model, with a half-duplex gate — when the agent is talking, your mic input is held (not uploaded), and when you interrupt, the agent's current sentence is dropped so the conversation feels natural. Your voice is processed through a single-use ephemeral token minted by the PetBuddy broker — no raw Google API key ever exists on your machine. The extension does not record or store conversations.
The tool set covers page actions (click, type, submit, scroll, scroll_to, get_page_map, read_page_text, read, fill_form), navigation (navigate, go_back, open_tab, switch_tab, list_tabs, close_tab), data (mine_data, extract_links, capture_page, save_data, screenshot), macros (start_recording, stop_recording, run_macro, list_macros, delete_macro), embodiment (perform animation, fly to position, hang_up, hide_buddy), and utilities (done, enter_map_mode, exit_map_mode, open_menu, menu_select, menu_run, research_note). Every tool is declared in a single registry, routed through the service worker, and guarded by the confirm gate for destructive actions.
No. Destructive actions — submit buttons, purchase buttons, delete actions — go through a deterministic confirm gate in the service worker. The gate requires a real human "yes" spoken within 15 seconds, on the same tab, with a label hash match (not the fragile marker number — word boundaries stop "pay" from firing inside "paypal"). The gate is code in the extension, not a prompt instruction the model can override. And there is a narrow corroboration rule: a "yes" spoken before a confirm appears is only spendable if the agent's own preceding sentence named that exact control. "Yeah, I like that shirt" must not press "Buy now."
The website embed ("one line of code") puts an AI agent on YOUR website for YOUR visitors. PetBuddy Anywhere puts the agent on EVERY website for YOU — it's your co-pilot, not your customer's concierge. The embed answers your customers' questions about your business. The extension answers YOUR questions about any page you're looking at, and operates it for you. Same underlying technology (Gemini Live, the broker, the voice pipeline), different direction: one faces your customers, the other faces you.
PetBuddy Anywhere is built and verified at v0.18.11: 390 offline checks, real-browser E2E 23/23, golden tasks 21/21, CI green on every push. It is pending Chrome Web Store review (unlisted — one $5 developer registration from Mario). In the meantime, you can install the website embed today: one line of code puts an AI agent on your own website right now, with the same voice quality, same character cast, and same Gemini Live backend. The embed is live on petbuddyconcierge.com — Vera is on this page. Tap her and ask.
Be clear on where the audio actually goes: your microphone audio and the text of any page you ask about are sent to Google and processed by Gemini Live to generate the replies. Our own server only issues a short-lived session credential for your licence — it never hears your audio and never sees the pages you are on. Voice is processed through a single-use ephemeral Gemini token — no raw API key ever touches your machine. The extension does not record, store, or transmit your browsing history. It does not run on websites you have not opened. You control when the microphone is active: tap to talk, tap to stop. The icon shows green when listening, white when idle. There is no "always-on" mode. The privacy policy is at petbuddyconcierge.com/privacy, and it discloses microphone use, audio processing, Gemini, Telegram (for lead delivery on the embed side), and that no personal information is sold.
PetBuddy AnywhereAn AI agent on every website. ~38 tools. Gemini Live voice.