Quay

Chat with the current page or let it browse for you. Bring your own key — OpenAI, Anthropic, Gemini — or go local with Ollama.

Quay is an open-source AI agent that lives in Chrome's side panel and actually drives your browser — it reads pages, moves an on-screen cursor, clicks, types, opens tabs, and finishes multi-step tasks while you watch. Every session is locked to its own tab group, so it never touches the rest of your browser. Bring your own key from any major provider — or run it 100% locally with Ollama. No subscription, no lock-in, zero telemetry. JUST TELL IT WHAT YOU WANT Plain English. Quay plans, navigates, and reports back inline — with screenshots, real markdown tables, and memory it can reuse later. • "Find the 3 best-rated sushi places near me and save their numbers." • "Fill this form with my info and submit it." • "Compare these 5 products and tell me which page loads cleanest." • "Pull every invoice from this dashboard into a table." WATCH IT WORK (new) Most agents act invisibly. Quay shows you everything in real time: • An on-screen cursor glides to each element and clicks — you see exactly what it's doing, where, and why. • A soft glow frames the page while the agent is in control, so you always know when Quay is driving. • A live action strip ("Clicking Sign in…", "Filling the form…") mirrors every step in the side panel. Transparent by design — no black box. BRING YOUR OWN KEY — ANY PROVIDER • OpenAI — GPT-4o, o-series • Anthropic — Claude Opus / Sonnet / Haiku • Google Gemini • OpenRouter — hundreds of models, pricing shown inline • Ollama — fully local, no cloud, no key Paste a key, click "Fetch models", pick one. Live token + cost counters on every message. BUILT TO ACTUALLY GET THINGS DONE • Per-tab-group sessions — research in one group, form-filling in another, in parallel, never crossing. • Vision-capable — screenshots are passed as real images, so the model sees what you see. • Batched actions & smart form-fill — Quay runs known flows in one shot and fills multiple fields by label in a single step. • Waits for the page — it knows when content has actually loaded instead of guessing, so it doesn't misfire on slow or dynamic sites. • Shared memory per session — saves facts, preferences, and resumable task state, and reuses them on later turns. • Recipes — record a successful flow once, then replay it, or schedule it (every N minutes, daily, weekly) to run headlessly and notify you. • Auto-compaction keeps long conversations inside the context window. • Web search built in (native or via Tavily) for live information. SAFE BY DEFAULT • Approval prompt before any action that changes a page, with a per-tool allowlist you control. • Plan-first mode — the model proposes a short plan before it acts. • Sensitive-site guardrails — on banking, government, health, and admin sites the allowlist is bypassed and every action asks first. • Prompt-injection hardening — page content is wrapped so the model treats it as data, never as instructions. • Login & CAPTCHA detection pauses the agent and hands control back to you. It never types passwords or solves CAPTCHAs. PRIVATE BY DESIGN • Your keys live in chrome.storage.local and never leave your machine except in direct HTTPS calls to the provider you chose. • Zero telemetry. Zero analytics. No outbound traffic other than the LLM calls you trigger. • Fully open source (MIT) — audit every line, fork it, self-host it. REQUIREMENTS Chrome 116+. An API key from any supported provider, or a local Ollama install for fully offline use. Source, issues & docs: https://github.com/nievesjesus/quay