Browser Agents in 2026: Browser Use vs Stagehand vs Skyvern vs Playwright MCP

Giving AI a browser stopped being one product category. It is at least four, sorted by who is driving, and in 2026 it acquired a fifth entrant that most comparisons miss: the browser you are already logged into. The frameworks converged technically — everyone grounds in page structure, everyone reaches for vision when structure fails — while diverging sharply in posture. Map your job to the posture and the choice mostly makes itself. Prices live on the tool pages so this page stays accurate between reviews.

Last reviewed: September 2026.

The summary table

The frameworks you build with

Browser Use — maximum autonomy

Browser Use is the category's breakout, at roughly 114,000 GitHub stars as of September 2026. You hand it a task and a model and it runs the perception-action loop itself. Around the MIT-licensed library there is now a full commercial platform: managed Chromium instances, hosted agents, a harness for evaluating runs, and MCP integration, with the API on V4 while V3 and V2 remain supported. Cloud usage is metered rather than sold in named tiers, and new accounts start with free credits.

Verdict: the right choice when the site is unfamiliar and the task is stated in a sentence. Its cost model is also its honesty — autonomy means a model call per step, and that is exactly what makes it the wrong shape for a flow you will run ten thousand times.

Stagehand — the engineer's posture

Stagehand is on v4, and the headline change since most write-ups is architectural: it drives the browser over the Chrome DevTools Protocol with no Playwright or Puppeteer dependency, while keeping Playwright-style APIs so the code still reads familiarly. Its three primitives — act(), extract() and observe() — let you write deterministic code everywhere and drop to AI only where selectors would rot, with schema-validated extraction. SDKs now cover TypeScript, Python and Go, and running on Browserbase adds server-side caching.

Verdict: the pick for automations a team owns and maintains. You give up the one-line convenience of an autonomous agent and get back something reviewable, testable, and cheap to run once the AI-assisted parts settle.

Skyvern — RPA replacement, not a developer tool

Skyvern aims at operations. It uses vision instead of brittle selectors, solves CAPTCHAs natively, and handles two-factor and TOTP login flows — the unglamorous essentials that real portal automation dies without. Workflows can be built by chat, by uploading a procedure document, by recording a browser session, or in a visual builder, with Python and TypeScript SDKs for code-first teams. It is AGPL-3.0, self-hostable with Docker, and available as a hosted product.

Verdict: choose it when the work is a documented business process against portals that fight back. The licence matters: AGPL-3.0 is a deliberate choice you should make consciously if you are embedding it in a product.

Tools for the agent you already have

The MCP servers are the right answer far more often than the frameworks admit, because they add capability without adding a runtime. Playwright MCP gives an agent structured browser automation from Playwright's accessibility tree rather than pixels, so no vision model is involved, across Chromium, Firefox and WebKit, with optional capabilities behind a flag for network mocking, storage, tracing, video, PDF generation and test assertions. Chrome DevTools MCP is the complement: Puppeteer-driven control of a real Chrome plus performance traces, network inspection and console access — note that it collects usage statistics by default and both that and update checks can be turned off.

Verdict: if you already live in a coding agent, start here. Playwright MCP to automate a flow, Chrome DevTools MCP to work out why the flow is broken. For agents verifying their own frontend work, this tier is unbeatable and costs nothing.

The browser you are already in

Claude in Chrome is the entrant that changes the shape of the category, and it is now generally available on paid Claude plans rather than a limited preview. It reads the page you are on and can click, type and fill forms without approving every step, while a separate safety check reviews each action for risk and for instructions hidden inside the site. Access is granted site by site through a permissions mode, sensitive actions such as purchases stop for explicit approval, and Team and Enterprise admins can enforce allowlists and blocklists org-wide.

Verdict: the pragmatic option for one-off work on sites you are already authenticated to, precisely because it inherits your session instead of rebuilding it. That is also the reason to keep it away from banking and health records, which Anthropic's own guidance says outright.

What you run it all on

Browserbase is the infrastructure layer: managed cloud browsers with stealth, proxies and live view, targeted by Stagehand natively and usable by anything speaking CDP. Local browsers are fine until you need concurrency, clean IP reputation, or a browser that survives your laptop closing.

When you do not need a browser agent at all

The cheapest browser agent is the one you did not run. If the task is reading the web rather than operating it, Firecrawl turns pages into markdown or JSON, handles JavaScript rendering and PDF or DOCX parsing, and offers search across the web returning full page content — with a free monthly credit allowance and paid plans above it. It also has an interact mode for content behind a click or a login, which covers a surprising share of what people reach for a full agent to do.

Verdict: try the API first. Escalate to an agent only when the task genuinely requires acting on state.

The security tax nobody escapes

Every option here reads hostile pages while holding a session, which makes prompt injection the category's shared tax rather than any product's flaw. The mitigations are architectural: domain allowlists, throwaway profiles carrying no valuable cookies, and human gates on payments, sends and deletions. Claude in Chrome ships this as product surface, and everyone else expects you to build it. The conceptual foundations — grounding, verification, and why an API beats a browser whenever one exists — are in How Computer-Use Agents Work, and the browser-agent-engineer agent encodes the same checklist for teams building on this stack.