One browser snapshot cost an AI agent 22,556 tokens, a developer measured

What the developer measured

As dev.to reported in a post by a developer who also sells a competing tool, letting an AI coding agent drive a browser can be surprisingly costly. Tools such as Claude Code, Codex and Cursor can control a web browser through a standard plug-in called MCP (Model Context Protocol). One popular option is Playwright MCP.

The author ran a single task on an admin dashboard holding 400 orders. The agent had to filter to the status Refunded, keep orders placed on or after 2026-08-01, and count the matches. He used Claude Sonnet inside Claude Code, took the median of 3 runs, and every run succeeded.

The result: one snapshot of the dashboard weighed 22,556 tokens. A token is the small chunk of text an AI model reads and bills by, so this is roughly the size of the page as the model sees it.

Why one snapshot gets paid for again and again

Playwright MCP does not send the agent a picture of the page. It sends an accessibility snapshot, which is a text outline of what is on the page, built for screen readers. That outline goes into the conversation with the model.

The catch is how AI models handle conversations. Each new call to the model re-reads everything said so far. So the snapshot is not charged once. It is charged on every call that follows.

The author's per-call numbers show this. Context stayed around the low twenties of thousands of tokens for the first few calls, then roughly doubled the moment the dashboard snapshot arrived, and stayed there for the rest of the task.

The author's fix and his numbers

The author built Deputy, a Chrome extension paired with a small server that runs on your own machine. Instead of handing the agent the whole page outline, Deputy turns the page's forms into typed tools. Each form becomes one call with named fields, plus a short list of things to click, so one call can fill a whole form.

On the same dashboard task and the same model, he reports these results:

• Agent turns: 10 with Playwright MCP, 3 with Deputy.
• Claude tokens: 245,250 versus 72,632.
• Page data sent to the agent: 46,640 characters versus 7,269.
• Total cost: $0.189 versus $0.072.

He also compared Deputy with Vercel's agent-browser, and all 12 runs passed. On the dashboard it took 11 turns and 271,172 tokens, against 3 turns and 72,756 for Deputy. On a 7-field booking form it was 7 turns and 155,200 tokens, against 3 and 68,683. He notes that part of agent-browser's count comes from a usage guide its own documentation tells the agent to load first.

Where the claim is weaker

To his credit, the author lists the cases where his tool does not win. On smaller pages the gap narrows: a booking form cost 107,392 tokens with Playwright MCP and 68,551 with Deputy, which is not a threefold difference. A short Wikipedia lookup came out about equal.

Deputy also depends on pages having real forms. Across 8 sites he tried, 7 produced typed tools on first load. On Wikipedia's appearance settings, the tool names still had to be guessed from field ids. Because clicks and keystrokes are simulated, heavy editors such as Google Docs may ignore them.

He adds that if you only test your own app in automated checks, Playwright itself is still the right tool. And if you stay with Playwright MCP, a setting called --snapshot-mode incremental can help.

What it costs and who benefits

Deputy has a free tier with 10 tools and no account key. Its Pro plan is $9.99 a month or $95.99 a year and unlocks all 27 tools. Pro also adds a hosted model that reads pages for your agent, up to 5,000 answers a month. The author says Pro can be cancelled any time, with a full refund within 14 days.

He also lists privacy points: the agent works in your own logged-in Chrome, anything that submits a form pauses to ask you on the page, and the local server listens only on 127.0.0.1, meaning your own computer.

Remember the author's disclosure that Deputy is his product. The numbers come from one task, one model and one author, though he says the method is public so others can rerun it. The most transferable lesson is the measurement itself: page snapshots are a major, repeating cost in browser agents.