Dowser vs Playwright MCP: your browser or a lab browser?

2 September 2026 · 10 min read · by Akhil Gorantala

Playwright MCP is Microsoft's open-source MCP server around the Playwright engine, and it deserves its popularity. If you are choosing between it and Dowser, you are really choosing between two different jobs, and this post is about being precise on which one you have. We build Dowser, so the bias is on the table; the claims below are stated so you can check them.

What Playwright MCP actually is

It launches a fresh browser the agent fully controls, and exposes tools for navigation, accessibility-tree snapshots, clicks, form filling and screenshots. Fresh browser is the key phrase: no cookies, no sessions, no extensions, no you. For tests and repeatable jobs a clean environment is a feature, and Playwright MCP delivers it across Chromium, Firefox and WebKit with the maturity you would expect from the Playwright project.

Where it hurts

Token burn. The model acts by reading page snapshots, and real-world pages produce huge ones. Comparisons of Playwright's own tooling put a typical MCP-driven task around 114,000 tokens versus about 27,000 for a CLI approach to the same task. That is per task, every time.

No logins. The lab browser starts signed out of everything. Anything behind your Gmail, your dashboards, your carts means scripting credentials into automation, which is exactly where most people stop.

Latency. Every action waits on a model round-trip to interpret the snapshot, so multi-step tasks accumulate seconds of pure waiting.

What Dowser does differently

Dowser lives in the browser you already use, as an extension paired with a local MCP server. Your assistant asks for "the add to cart button" and a local ranker resolves it against the page's accessibility tree and DOM in a few milliseconds, with no model call and no page content leaving your machine. Your logins are just there, because it is your browser. You watch every action happen with a visible cursor, moving your mouse pauses it instantly, and password or card fields are refused outright.

Side by side

The token math, worked through

Take the linked comparison's numbers at face value: roughly 114,000 tokens for a typical MCP-snapshot task. If your assistant touches the browser ten times a working day, that is over a million tokens a day spent on page snapshots, before any of the work you actually wanted the model to do. At typical frontier-model prices that is real money per month, but the quieter cost is context: those snapshots sit in the conversation window, crowding out your code and your instructions, which is why long browser-heavy sessions with snapshot tools get worse as they go. A Dowser action is a sentence-sized tool call; a hundred of them cost less context than one snapshot. And no, a cheaper model does not fix it: the dollars shrink but the snapshot still has to be produced, shipped and read on every action, so the latency and the context crowding stay.

Setting them up

Playwright MCP: claude mcp add playwright -- npx @playwright/mcp@latest (or the equivalent in Cursor and friends), and it downloads its browser on first run.

Dowser: paste one line into your assistant: Set up the Dowser browser extension for me: fetch https://dowser.click/install.md and follow the steps in it, then confirm Dowser is connected. You press Add to Chrome once; the 5-day trial starts on its own. The full walkthrough with guardrails explained is in the Claude Code guide.

Which to pick

Playwright MCP when: you are writing automated tests or CI jobs that must be reproducible, the task should run isolated from your real sessions, or you need Firefox and WebKit coverage.

Dowser when: the task lives in your day-to-day browser with your accounts and state, you do this many times a day so speed and token cost compound, and you want the assistant you already use to do it rather than a separate script. The wider field, including Browser MCP and Nanobrowser, is mapped in the full comparison.

FAQ

Can Playwright MCP use my existing Chrome profile?

It can be pointed at a persistent profile, but that fights its design: the tool is built around clean, reproducible sessions, and mixing in your live profile gives up exactly the isolation it is good at.

Does Dowser replace my Playwright tests?

No. Dowser has no test runner, no assertions and no CI story; it is for an assistant doing live tasks in your browser, not for a suite you run on every commit.

Is the token gap smaller with a cheaper model?

The dollars shrink, the latency and the context crowding do not: the snapshot still has to be produced, shipped and read on every action, whatever model reads it.

Can they run at the same time?

Yes, they are independent MCP servers driving different browsers. A common setup is Playwright MCP in CI and Dowser in the working browser.

Does Dowser work outside Chromium?

No: Chrome, Brave, Edge and other Chromium-based browsers. Firefox and Safari are not supported, and if you need them automated, Playwright is the right tool.

What happens to Dowser when the page has no accessibility tree?

Canvas-drawn interfaces with no accessible structure are its weak spot; it will return low confidence rather than guess. Screenshot tools handle those pages better, which is part of why some people run both.