Dowser vs Browser MCP: same idea, different engine

2 September 2026 · 10 min read · by Akhil Gorantala

Browser MCP got an important thing right early: AI browser automation belongs in the browser you already use, with your logins and sessions, not in some headless clean room. Dowser agrees, which is why the two look similar from a distance. This post, written by the people who build Dowser, is about the one place they genuinely differ: the engine underneath each click.

Where they are the same

• Both are a Chrome extension plus a local MCP server your assistant connects to.
• Both work with any MCP client: Claude Code, Cursor, Windsurf, VS Code and friends.
• Both act in your real, logged-in browser, so sites see you and not a bot farm.
• Both keep the loop local rather than routing your browsing through a cloud.

That shared premise matters: it is the pro-user, pro-privacy end of the category, and either choice beats shipping your browsing to a datacenter. The rest of this post is about what happens after the premise.

The engine difference

When your assistant wants to click something through Browser MCP, it requests a snapshot of the page, reads it with the model, decides which element to target, and issues the action. The model is inside the loop for every step. That is flexible, and it is also where the time and the tokens go: big pages make big snapshots, and the model has to process one per action.

Dowser moves that decision out of the model. The assistant states the intent once, in plain English: find "the newsletter unsubscribe link". A ranker inside the extension scores the page's real elements, from the accessibility tree and DOM, and returns the match in a few milliseconds with a confidence score. High confidence acts immediately; low confidence returns candidates instead of guessing. No page snapshot ever enters your model's context window just to find a button.

Side by side

What it means in practice

Speed: a find or click resolves in single-digit milliseconds instead of a model round-trip, so a five-step task finishes before a snapshot tool completes its first step. Token cost: a three-click task is three short tool calls, not three page dumps, which also means long browser sessions do not crowd your context window. Privacy: page content is processed on your machine and stays there. Guardrails: a visible cursor for every action, instant handoff when you move the mouse, refusal of password and card fields, and confirmation required for destructive clicks. Form work shows the difference most clearly; the worked example is in the form-filling guide.

How the timing was measured

The millisecond claims are our own runs: we time from the moment the MCP server receives the find or act request to the moment the click has fired in the page, on a real product page, on ordinary hardware. Typical results are single-digit milliseconds for a find and a click. We publish them as typical figures from our own runs, not vendor benchmarks, and the homepage demo is a screen recording of one such run rather than an animation. For any snapshot-driven tool the equivalent interval necessarily includes a model round-trip, so it is bounded below by your model's latency, usually several hundred milliseconds at best.

Setup, side by side

Browser MCP: install the extension, add the server with claude mcp add browsermcp -- npx @browsermcp/mcp, then connect the extension to the tab you want automated.

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. It downloads the local piece, pairs it, and registers itself; you click Add to Chrome once. Step-by-step version with screenshots of the flow: the Claude Code guide.

The maintenance problem

This section is uncomfortable to write about a tool we respect, but it is the practical heart of the comparison in late 2026: Browser MCP's repository has had no commits since April 2025, roughly seventeen months, while about 150 issues have accumulated unanswered, and the extension's last Chrome Web Store update was May 2025. A hundred thousand people installed a genuinely good idea whose maintenance stopped. That matters concretely: MCP clients evolve, Chrome ships breaking extension changes, and sites change; an unmaintained bridge decays even if nothing looks broken today. Dowser is a paid product precisely so that maintaining it is someone's job; it shipped fixes this week.

Switching takes about two minutes

If you are a Browser MCP user weighing a move, the shapes match: extension plus local MCP server, your real browser, any MCP client. The migration is: remove the browsermcp entry from your MCP config, paste the one-line setup 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.), and click Add to Chrome once. Your assistant's prompts do not change; the tools it sees become dowser_find, dowser_act, dowser_read, dowser_navigate and dowser_tabs, and the 5-day trial gives you a real overlap period before anything is paid.

Price, honestly

Browser MCP is free, and free is a real feature. Dowser is $19 a month after a 5-day trial, and the money buys the local ranking engine, the safety layer and support from a developer whose product this is. If your assistant touches the browser a few times a week, free wins. If it is part of how you work every day, milliseconds and quiet token bills compound fast, and the arithmetic is worked through in the Playwright MCP post, whose token numbers apply to any snapshot-driven engine.

FAQ

Can I run both at once?

Yes, they are independent extensions and servers. Some people keep Browser MCP for snapshot-style page reading and use Dowser for the clicking; nothing conflicts as long as they do not act on the same tab simultaneously.

Does Dowser ever fall back to a model?

No. If the ranker is not confident it returns the candidates to your assistant and lets it choose, which costs one small tool response rather than a snapshot.

What about pages that change while you act?

Both tools act on the live DOM, so a page that rerenders mid-task affects both equally; Dowser re-runs its find in milliseconds, which makes retrying cheap.

Which reads pages better?

Different jobs. Browser MCP's snapshots give the model a full-page view in one shot; Dowser's dowser_read returns a compact outline or readable text sized for your context window. For "read this article", both work; for "how is this page structured", the snapshot is richer and costs more.

Do sites detect either as a bot?

Both act inside your real browser session, so there is no separate fingerprint. Behavior is what a site sees, and both tools click like a fast human rather than a datacenter farm.

Is Browser MCP open source?

Its server and protocol pieces are public and free to use; check its repository for current license details. Dowser is commercial with a tracked public build for review on the Chrome Web Store.