Claude in Chrome vs Dowser: screenshots or milliseconds

2 September 2026 · 10 min read · by Akhil Gorantala

Claude in Chrome is Anthropic's own extension, and if you searched for an alternative to it you probably hit one of its three edges: the pause before every click, the plan requirement, or wanting the same hands for assistants that are not Claude. Dowser is built against exactly those edges, so this comparison writes itself; we build Dowser, so check the claims against the recorded demo on the homepage.

How each one clicks a button

Claude in Chrome is computer use in your browser: it captures a screenshot, Claude's vision reads it, decides coordinates, and a virtual cursor moves and clicks. Every action is a round-trip through a vision model, which in our runs lands around 1,300 ms per find-and-click, and each screenshot of the page travels to Claude.

Dowser never renders the page to pixels. Your assistant names what it wants ("the add to cart button") and a ranker in the extension resolves it against the page's accessibility tree and DOM, locally, in about 6 ms in the same runs, then clicks with a visible cursor. The timing methodology is the same for both: tool request to completed click, same page, same machine, typical figures rather than a benchmark suite.

Side by side

Where Anthropic's extension is genuinely better

If your browsing leans on visually unusual interfaces, keep it; the vision loop is the right tool there, and running it alongside Dowser costs nothing but an extension slot.

Where Dowser wins

Frequency and privacy. At ten browser touches a day, the difference between 1,300 ms and 6 ms per action is the difference between waiting on your tools and not noticing them. Nothing about the page leaves your machine, which matters on logged-in dashboards, email and anything with customer data on screen. And the same hands serve Cursor, Codex and Gemini CLI, so switching assistants does not mean losing browser control. The wider three-way comparison including Browser MCP is in the Claude browser automation guide.

The numbers over a working day

Make it concrete. A modest day of assistant browsing is 30 actions. Through a screenshot loop at ~1,300 ms that is around 39 seconds of pure waiting, spread as pauses that break your flow, plus 30 screenshots' worth of vision tokens from your plan. Through a local ranker it is under a fifth of a second of waiting all day and zero find-related tokens. Scale to a heavy day of 150 actions and the gap is three minutes of waiting versus under one second. None of this makes the vision approach wrong; it makes it expensive per action, and per-action costs are the ones that compound.

Running both, and setup

They coexist cleanly: some people keep Claude in Chrome for canvas-heavy pages and let Dowser handle everything with real DOM structure, which is most of the web. Claude in Chrome installs from the Chrome Web Store and signs into your Claude account. Dowser sets itself up from one pasted line:

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.

Two minutes, one Add to Chrome click from you, 5-day trial, no card. Step-by-step version: give Claude Code hands in your browser.

FAQ

Can I run both?

Yes. They are separate extensions; just avoid pointing both at the same tab at the same moment.

Does Dowser need a paid Claude plan?

No. Any MCP assistant works, including free tiers; the license is a one-time $29 for the extension itself after the 5-day trial, with a 14-day refund.

Is 6 ms marketing math?

It is a typical measured run, not a guarantee; complex pages can take tens of milliseconds. The structural claim is the honest one: no network and no model in the find path, so it cannot be slower than the page itself allows.

What happens on a page Dowser cannot parse?

It returns low-confidence candidates or nothing, and your assistant tells you. That is the moment to use Claude in Chrome's vision, or your own hands; Dowser never guesses blind.

Why is Claude in Chrome slower if Anthropic builds the model?

Physics, not engineering quality: capturing a screenshot, uploading it, running vision inference and returning coordinates is a network round-trip through a large model. A local ranker skips the network and the model entirely, which is the whole design difference.

Which should a team standardize on?

Mixed-assistant teams (some Cursor, some Claude Code) get more from MCP-based hands, since one tool serves everyone; all-Claude teams can reasonably run the official extension plus Dowser for the daily-driver speed.