Paw Work Makes the AI Ask What You’re Pointing At

The Ritual of Selection

In the Chinese tradition of Zhuazhou, parents place objects before a child on their first birthday and wait to see which one the child grabs. The choice is interpreted as a sign of future inclinations—a book for scholarship, an abacus for commerce, a seal for bureaucracy. The ritual is bounded, intimate, and entirely about the objects within arm’s reach. There is no expectation that the child will crawl across the room to find something else. Contemporary photographers still document the ceremony, noting how children return again and again to certain objects—a steamed bun, a gavel, a book—while parents laugh and read meaning into the choice.

Paw Work, a Chrome Manifest V3 extension recently circulating in open-source circles, borrows this logic in both name and spirit. The project’s Chinese title is 爪爪—literally “paw paw”—and its interaction model is deliberately constrained: you turn on Paw Mode, click the images, tables, text blocks, or links on the page you are already viewing, and tell the sidepanel what you want. The agent does not roam. It does not open new tabs, fill checkout forms, or submit expenses while you sleep. It inspects only what you have pointed to, then returns an editable artifact—a spreadsheet in Univer, a slide deck on a tldraw canvas, a real HTML site, or a document. The metaphor is almost too perfect: like the Zhuazhou child, the agent is given a limited set of objects and asked to make something of them.

The Anti-Agent Stance

This boundedness is the project’s most deliberate feature, and it arrives at a moment when the industry is sprinting in the opposite direction. The AI browser-agent market is projected to grow from $4.5 billion in 2024 to $76.8 billion by 2034, and GitHub repositories like Browser Use and Firecrawl have accumulated tens of thousands of stars by promising autonomous navigation—agents that accept a natural-language goal, click through pages, fill forms, extract data, and complete multi-step workflows without human intervention. The taxonomy now distinguishes between AI-assisted browsers (Edge Copilot, Brave Leo) and fully agentic systems (Comet, Atlas, Skyvern, Stagehand) that “take an outcome and figure out the steps.”

Paw Work’s documentation explicitly rejects this entire framing. “It is not an unattended Operator that roams the web for you, and not a terminal coding agent,” the README states. The tool is designed for “people who already work in the browser—compare, extract, export, turn a page into a sheet, deck, poster, site, or long doc.” In other words, it assumes you are already logged in, already looking at the SaaS dashboard or product listing or research article, and simply want to transform that visible material into a structured file. The agent’s job is not to discover; it is to redescribe.

This is a meaningful philosophical split. Autonomous agents trade reliability for convenience: task success rates on benchmarks like WebVoyager range from roughly 30 percent to 89 percent depending on the framework, and the probabilistic nature of multi-step navigation means a failed run can leave accounts, carts, or databases in unpredictable states. Paw Work sidesteps this by never acting outside the selection. It cannot click “purchase” because it does not click at all. It reads what you selected, generates a plan card if the task is complex, and waits for approval before proceeding. The user remains the operator; the extension is a very smart pair of hands.

Artifacts, Not Essays

The second displacement is in the output. Most AI browser tools, even the assisted ones, return text: summaries, answers, code snippets in a chat panel. Some sidebar add-ons, like those populating the Google Workspace Marketplace, promise to bring generative AI “directly into your document—no more switching tabs or copying and pasting,” yet they still largely annotate or rewrite content inside an existing editor. Paw Work returns files on live canvases. Point at a set of product photos and ask for a comparison sheet, and you get a live Univer spreadsheet you can keep editing. Select notes and request an eight-page deck, and you receive tldraw slides that export to PPTX. Ask for an editable site and you get real HTML, not a static screenshot or a React component buried in a chat thread.

This matters because the boundary between “AI output” and “working document” has become a friction point in knowledge work. When a model returns a markdown table in a chat window, the user must copy, paste, and reformat. Paw Work attempts to collapse that gap by generating directly into embedded office suites. The spreadsheet is not described; it is the spreadsheet. The slide is not suggested; it is on the canvas. Even the generated JavaScript runs locally inside a QuickJS VM compiled to WebAssembly—no CDN executable code, no access to Chrome extension APIs, no visibility into the live page DOM. The sandbox is as strict as the selection model.

Bring Your Own Everything

Paw Work also arrives amid a broader shift toward “bring your own key” infrastructure. Microsoft’s VS Code recently expanded its BYOK support, allowing users to plug in their own API keys for OpenAI-compatible endpoints, Ollama, OpenRouter, and others rather than relying on a centralized Copilot quota. Paw Work extends this ethos to browser productivity: no account, no Paw Work server, no hosted model tier. You configure your chat endpoint, image generation key, and web-search key in the sidepanel settings. Credentials live in Chrome extension storage on your machine and are sent only to the HTTPS endpoints you configure.

This is not merely a privacy posture; it is an architectural one. Because there is no backend, there is no vendor lock-in, no rate-limit negotiation, and no surprise pricing tier. The extension is a local client that happens to speak LLM protocols. For users with existing API access—whether through OpenRouter, a self-hosted model, or a corporate endpoint—this means the tool is pure margin. For users without keys, it means the tool does nothing at all, which the documentation states bluntly: “Nothing runs without them.”

Where It Sits in the Agent Zoo

Classifying Paw Work using the current browser-agent taxonomy is instructive. It is not a consumer AI browser like Perplexity Comet or ChatGPT Atlas, because it does not replace navigation or search. It is not an open-source agent framework like Browser Use or Stagehand, because it does not expose a programmable API for autonomous task completion. And it is certainly not managed infrastructure like Firecrawl or Browserbase, because there is no cloud session rental, no per-page credit system, no remote browser sandbox.

It is closer to a scoped productivity layer—an extension that treats the current tab as the universe of discourse. In this sense it competes less with the agentic frameworks and more with the glut of sidebar chat extensions that summarize or rewrite content inside Docs. But even there, the difference is sharp: those tools typically annotate or augment the page you are on, whereas Paw Work extracts from it and generates a new file in a parallel workspace.

Rough Edges and Open Questions

The project is not yet listed on the Chrome Web Store, so distribution currently requires sideloading the unpacked branch. The documentation is admirably clear about this, repeatedly warning passersby not to clone the source repository itself, which includes hundreds of megabytes of Node modules. The Web Store is described as the “later front door.”

There are material limitations. The tldraw canvas shows a watermark unless you supply a separate production license. Capture is intent-based, not a full DOM dump, which means the agent may need to ask clarifying questions when the outcome type is ambiguous. And the tool is explicitly “not a full Excel / PowerPoint / browser-RPA replacement.” It will not unattendedly roam, pay, or submit forms.

Perhaps the most interesting tension is whether the Zhuazhou model scales. Autonomous agents are unreliable precisely because the open web is messy; Paw Work solves this by refusing to engage with the mess beyond the user’s selection. But as web applications grow more dynamic—single-page apps with infinite scroll, nested shadow DOM, authenticated data that requires interaction to reveal—the simplicity of “click what matters” may become a bottleneck. The agent can inspect on demand, yet it cannot interact with the page on your behalf to surface hidden data. Whether that constraint is a feature or a bug depends entirely on whether you believe the future of AI assistance is scoped augmentation or autonomous delegation.

Outlook

Paw Work suggests a third path in a landscape bifurcated between chatbots and roving agents. It treats the browser not as a territory to be conquered but as a workspace to be harvested. The user points; the agent interprets; the file appears. There is something almost old-fashioned in this contract: the human sets the scope, the machine executes within it, and the output is a tangible artifact rather than a conversational promise.

Whether that contract proves durable will depend on how many everyday workflows genuinely fit the “select and redescribe” pattern, and whether users prefer the certainty of a short leash to the seductive chaos of an agent let loose. For now, Paw Work is a reminder that not every AI tool needs to be an operator. Sometimes it is enough to be the ritual interpreter of the objects already on the table.