Page Agent

Page Agent (alibaba/page-agent) is an open-source JavaScript agent that lives inside your web page and operates its interface from natural language commands — no browser extension, Python backend, or headless browser required. You add one script tag (or npm-install the TypeScript package) and the agent reads the page's DOM as text, then clicks, types, and navigates on the user's behalf. Because it is text-based DOM manipulation rather than screenshots, it works with ordinary (non-multimodal) LLMs, including locally deployed models — you bring your own LLM via any OpenAI-compatible endpoint, and a free demo LLM CDN build is available for technical evaluation. Real features include a one-line CDN integration, a programmatic API (new PageAgent({model, baseURL, apiKey}) then agent.execute('Click the login button')), an optional Chrome extension for multi-page/multi-tab tasks, and a beta MCP server that lets external agent clients (like Claude Desktop or other MCP hosts) control the browser from outside — a notable differentiator for MCP-based workflows. Use cases called out by the project: shipping a SaaS AI copilot in a few lines of code, smart form filling that turns 20-click ERP/CRM workflows into one sentence, and accessibility via voice or natural-language control of any web app. It builds on browser-use's DOM processing, is MIT licensed, free, and has 26k+ GitHub stars with an active docs site and Chrome Web Store extension.

What Page Agent does in practice

One script tag adds an AI agent to any web page — no extension or headless browser. Text-based DOM understanding: works without screenshots or multimodal LLMs. Bring your own LLM: mainstream APIs or locally deployed models. MCP server (beta) lets external MCP clients control the browser. Chrome extension for multi-page, cross-tab agent tasks. TypeScript, MIT licensed, small bundle published on npm. These capabilities should be judged as a workflow, not isolated checkboxes. Concrete best-fit scenarios include: Embedding an AI copilot into a SaaS product without a backend rewrite. Natural-language form filling in ERP, CRM, and admin systems. Accessibility: voice/natural-language control of web apps. Extending an existing web agent across browser tabs. Letting MCP agent clients (e.g., Claude Desktop) operate live web pages.

Pricing and total cost

Open Source is listed at Free; MIT licensed npm package and CDN build; Free demo/testing LLM API for evaluation; Bring your own LLM (OpenAI-compatible, including local models); Optional free Chrome extension for multi-page tasks. These figures come from the staging record and vendor pages were fetched during this revision, but pricing is mutable. Recheck the live pricing page, checkout, or quote before publishing or buying. Include required seats, usage credits, model or API charges, storage, implementation, support, taxes, and overages. A free tier can prove compatibility; it does not necessarily model production volume or support needs.

Strengths and tradeoffs

Practical strengths include: One script tag adds an AI agent to any web page — no extension or headless browser. This directly supports the primary workflow. Text-based DOM understanding: works without screenshots or multimodal LLMs. Bring your own LLM: mainstream APIs or locally deployed models. MCP server (beta) lets external MCP clients control the browser. Important constraints include: Plan prices can omit usage overages, model-provider charges, implementation work, taxes, or required seats. Features and limits can change by plan, billing interval, region, and release; verify current documentation. MCP connectivity expands the permission surface and requires narrow scopes, approvals, logs, and credential rotation. A representative pilot is necessary because a feature list does not establish reliability or operational fit. A capable product can still be the wrong operational choice. Document who reviews outputs, who owns failures, what data may enter the system, and how work continues during an outage.

How to evaluate Page Agent

Use a two-week proof of concept on a noncritical repository or workflow. Define four repeatable tasks: one small change, one multi-file change, one failure-recovery scenario, and one permissions-sensitive action. Measure setup time, completion rate, human review minutes, retries, API or token cost, and defects found after acceptance. Pin versions, inspect generated changes, and test rollback. Give the tool only the repositories, APIs, and scopes needed for the trial. Verify export, backup, logging, rate-limit, and offboarding behavior before treating a fast demo as production readiness.

Bottom line

Page Agent is worth shortlisting when its capabilities match a measured workflow and the buyer can verify current limits. Do not select it from a feature list alone. Compare at least two alternatives using the same inputs, acceptance criteria, and cost model. Preserve test results so renewal is based on outcomes rather than adoption anecdotes.