A browser agent that returns files, not chat essays

What it does

Paw Work is a Chrome MV3 extension that sits beside your normal browsing session. Flip on Paw Mode, click to select images, tables, text, or links on the page you already have open, then describe what you want in the sidepanel. It hands back editable files—Univer spreadsheets, tldraw slides that export to PPTX, real HTML sites, or documents—instead of dumping a wall of text.

The interesting bit

The sandboxing model is unusually paranoid in a good way: generated code compiles with an embedded esbuild-wasm and runs inside a QuickJS VM, so it never sees chrome.* APIs, the live DOM, or another session’s files. For complex tasks, it can emit a plan card first and wait for your approval before touching the canvas, so it does not silently expand beyond what you selected.

Key highlights

• BYOK and no server: you supply the LLM keys, which stay in Chrome’s local extension storage; there is no Paw Work cloud or account
• Selection-first scope: it captures only what you point at, not the whole site, and inspects on demand
• Live canvases: outputs open in embedded Univer (sheets), tldraw (slides/posters), or a real HTML workspace, all persisted locally via IndexedDB/OPFS
• Plan cards: for multi-step work, approve, decline, or leave notes to revise the plan before execution
• Isolated runtime: model-generated JavaScript runs in a WASM sandbox, not your main browser context

Caveats

• Not on the Chrome Web Store yet; you must sideload the unpacked branch manually
• Requires your own API keys for chat, image generation, and web search/fetch
• tldraw’s production watermark appears on Design/Slides unless you bring a separate tldraw license
• Explicitly not an unattended RPA: it will not submit forms, check out, or pay on your behalf

Verdict

Worth a look if you spend your day comparing products, extracting tables, or turning articles into decks and want to stay inside the browser. Avoid if you need an autonomous operator that roams across sites to complete transactions, or a terminal-centric coding agent.