Chrome DevTools MCP for agents: 393,149 tokens per Wikipedia page

Chrome DevTools MCP gives an agent the same browser a web developer debugs in, and it is a good tool for that job. Operating pages for hours is a different job, and the difference shows up in tokens. We ran the same action script through Chrome DevTools MCP 1.9.0 and five other browser tools on real pages. Seeing one Wikipedia article cost 393,149 tokens, and a 1 million token context window was full after 2 steps on that page.

What is Chrome DevTools MCP?

Chrome DevTools MCP is the Model Context Protocol server of the Chrome DevTools team, and its repository describes it as a server that "lets your coding agent (such as Antigravity, Claude, Cursor or Copilot) control and inspect a live Chrome browser". It "uses puppeteer to automate actions in Chrome", and its tools cover input, navigation, emulation, performance traces, network requests and debugging. The same README carries a warning worth reading before you point it at a signed-in browser: it "exposes content of the browser instance to the MCP clients allowing them to inspect, debug, and modify any data in the browser or DevTools."

How many tokens does it spend?

Chrome DevTools MCP spent between 10,346 and 393,149 tokens to let an agent see a page, and a follow-up step cost almost as much as the first look. A step is one action plus everything the agent needs to act next, so Chrome DevTools MCP got an explicit snapshot call after each action, the call an agent needs before it can pick the next element. Tokens were counted with OpenAI's tiktoken and the cl100k_base encoding, so the absolute numbers approximate any model's tokenizer and the ratios hold for every BPE.

The second number in each row is the one that decides the bill. Chrome DevTools MCP answers each step with the whole snapshot again, so a follow-up step on Hacker News costs 13,311 tokens, nearly the 13,331 of the first look. On the Wikipedia article one answer is so large that a 1 million token window holds only 2 of them.

Why does a whole-page answer cost so much?

A whole-page answer costs so much because the model API is stateless: every step resends the whole conversation, so each snapshot is paid again at every later step. The accessibility tree behind the snapshot is large to begin with. Playwright alone, without any MCP server, reads a raw accessibility tree of 3,753,843 tokens from the same Wikipedia article, and every snapshot tool starts from that tree and differs only in how hard it filters. Chrome DevTools MCP filters it to 393,149; a map of only the elements an agent can operate, sent as a difference after each action, stays at 2,771 for the first look.

The tool catalogue is not where the money goes. The 29 tool definitions cost 4,849 tokens on every request, a fixed fee that comes with all MCP servers for browsers, and it is smaller than the 21 tools Scalebrowser ships by default. The page data is the cost.

How does Playwright MCP compare?

Playwright MCP 0.0.80 costs more for the first look and about the same after it. It spent 604,785 tokens to see the Wikipedia article, but 12,494 per follow-up step on Hacker News against 13,311 here, and $21.06 for 25 steps against $22.50. It also writes large snapshots to a file whose content the agent still has to read, and our count of Playwright MCP tokens includes that content. Both servers were built around a developer's view of the page: the full tree, every time.

Should an agent use Chrome DevTools MCP or Playwright MCP?

Use Chrome DevTools MCP when the agent's job is to debug a page, because performance traces, network inspection and console access are what it is built for. On a small page for a short session the token cost does not matter much: the GitHub repository cost 10,346 tokens to see. For long sessions on large pages, both servers run out of context window long before the task is done, and the step count is the number to check first. The MCP server page of our documentation explains how a page map answers with only what changed, and the tool reference lists what an agent can call.