---
title: "Browser MCP servers compared: tokens per step across 6 tools"
description: "Six browser MCP servers measured on live pages: tokens to see a page, tokens per step, the tool list's fixed cost, and which server fits which job."
canonical: "https://scalebrowser.net/blog/browser-mcp"
last_modified: "2026-09-25"
published: "2026-09-25"
author: "davide"
category: "infrastructure"
---
> ## Documentation Index
> Fetch the complete documentation index at: https://scalebrowser.net/llms.txt
> Use this file to discover all available pages before exploring further.
# Browser MCP servers compared: tokens per step across 6 tools
> Six browser MCP servers measured on live pages: tokens to see a page, tokens per step, the tool list's fixed cost, and which server fits which job.
A browser MCP server is how most agents reach a web page today, and it decides two things at once: what the agent can do, and what every step costs. We ran the same action script through six of them on four live pages and counted every answer in tokens. The spread is two orders of magnitude, and on a large page four of the six fill a 1 million token context window within 4 steps.
## What is a browser MCP server?
A browser MCP server is a Model Context Protocol server whose tools open pages, read them and act on them for a language model. The [MCP specification](https://modelcontextprotocol.io/specification/2025-06-18/server/tools) defines tools as functions a server exposes "that can be invoked by language models", discovered through a `tools/list` request in which each tool carries a name, a description and an `inputSchema`. A browser server adds two decisions of its own on top: how it shows the page to the model, and how much of it comes back after each action.
## How many tokens does each browser MCP server spend per step?
A step here is one action plus everything the agent needs for the next one, because a tool whose click answer is cheap but leaves the agent unable to act has only moved the cost to a second call. Every server got exactly the calls it needs to stay able to act, and files that a server writes instead of answering were read and counted, because the agent has to read them too. Tokens were counted with OpenAI's [tiktoken](https://github.com/openai/tiktoken) and the `cl100k_base` encoding; the ratios between servers hold for any BPE tokenizer.
| Server | Tokens to see the page | Tokens per next step |
| --- | --- | --- |
| Scalebrowser | 4,309 | 462 |
| camofox-browser | 12,112 | 12,076 |
| Playwright MCP 0.0.80 | 24,990 | 12,494 |
| Chrome DevTools MCP 1.9.0 | 13,331 | 13,311 |
| browser-use 0.13.10 | 18,097 | 18,087 |
| stealth-browser-mcp | 26,478 | 26,438 |
Only one of the six answers after an action with the difference since the last reading; the other five send the page again, so their next step costs about as much as their first look. On a large page the difference decides whether a session exists at all.
| Server | Tokens to see the page | Steps until the window is full |
| --- | --- | --- |
| Scalebrowser | 2,771 | over 400 |
| browser-use 0.13.10 | 8,369 | 100 |
| camofox-browser | 205,936 | 4 |
| Playwright MCP 0.0.80 | 604,785 | 3 |
| Chrome DevTools MCP 1.9.0 | 393,149 | 2 |
| stealth-browser-mcp | 1,057,189 | 0 |
stealth-browser-mcp returns more than a million tokens for this one page, so the session ends before its first step can be paid. Three of the six servers write large pages to a file and return a path or a page of it instead; counting only the answer would have measured 300 tokens for stealth-browser-mcp instead of 1,057,189.
## Why does the tool list cost tokens on every request?
The tool list costs tokens on every request because the model API is stateless: the tool definitions travel with every call to the model, including the calls that use no tool. That makes the catalogue a fixed fee per request, and its size is set by the number of tools and the length of their descriptions.
| Server | Tools | Catalogue tokens per request |
| --- | --- | --- |
| stealth-browser-mcp | 97 | 12,590 |
| Scalebrowser, as shipped | 21 | 5,267 |
| Chrome DevTools MCP 1.9.0 | 29 | 4,849 |
| Playwright MCP 0.0.80 | 24 | 3,162 |
| browser-use 0.13.10 | 16 | 1,036 |
| camofox-browser | 11 | 932 |
On Hacker News the shipped Scalebrowser catalogue is 21 percent of the input over 25 steps and 11 percent over 100. On Playwright MCP the catalogue is 2 percent, not because it is small but because the page data next to it is so large. How many tools a server should expose follows from that fee: as many as the task needs by default, the rest on request. Scalebrowser ships 21 of its 68 tools in the default catalogue for that reason, because every extra definition is paid again on every request of every session.
### What does Playwright MCP cost?
[Playwright MCP](https://github.com/microsoft/playwright-mcp) 0.0.80 spent 604,785 tokens to see the Wikipedia article and 12,494 per step on Hacker News. Since the 0.0.8x releases it writes the snapshot to a file and returns only the path: 86 tokens at the answer, 301,197 once the agent reads the file. The case study on [Playwright MCP token usage](/blog/playwright-mcp-tokens) prices a 75 step session.
### What does Chrome DevTools MCP cost?
[Chrome DevTools MCP](https://github.com/ChromeDevTools/chrome-devtools-mcp) 1.9.0 spent 393,149 tokens to see the Wikipedia article and 13,311 per step on Hacker News, where 25 steps cost $22.50 at a list price of $5 per million input tokens. Its README describes it as a way to let "your coding agent" control and inspect Chrome, and its tools for performance traces and network requests fit that job. Our [Chrome DevTools MCP](/blog/chrome-devtools-mcp) review weighs it against Playwright MCP for agent work.
### Why does a page map beat raw HTML?
A page map beats raw HTML because it lists only what an agent can operate, each element with its role, its name and a reference. The raw material is huge: without any MCP server, Playwright reads the Wikipedia article as 985,710 tokens of HTML and 3,753,843 tokens of raw accessibility tree, and every snapshot tool starts from that tree and differs only in how hard it filters. To check that a small map is not small by omission, we added the page's full wording through a separate read call: Wikipedia then costs 5,688 tokens, and the Hacker News map still carries 391 addressable elements, every story link and every comment link among them.
## So which browser MCP server should you pick?
| Situation | Pick | Why |
| --- | --- | --- |
| Debugging a page you are building | Chrome DevTools MCP | performance traces, network and console are its job |
| Short test sessions on small pages | Playwright MCP | a full accessibility snapshot per answer; example.com costs 269 tokens to see |
| An open-source agent loop you want to own | [browser-use](https://github.com/browser-use/browser-use) | the loop is its product; about 18,000 tokens per step on Hacker News |
| A Firefox build behind a REST server | [camofox-browser](https://github.com/jo-inc/camofox-browser) | 11 tools and the smallest catalogue, but 205,936 tokens for a large page |
| Long agent sessions on large pages | a server that answers with the difference | 462 tokens per step keeps a session going for 1,484 steps on Hacker News |
Our verdict: for an agent that operates pages over many steps, the answer format decides the cost, and the tool count comes second. Pick a server that answers with what changed, and keep its default catalogue small. [stealth-browser-mcp](https://github.com/vibheksoni/stealth-browser-mcp) is the one we would not use for agent work on large pages, because its first answer on the Wikipedia article did not fit into a 1 million token window.
### Which browser does the MCP server drive?
The MCP server is only the interface; the browser underneath decides whether the page lets the agent in at all. The guide to [browsers built for AI agents](/blog/ai-agent-browser) walks through the four layers a site reads. The [MCP server page of our documentation](/docs/agents/mcp-server) describes how Scalebrowser's page map and difference work, and [the tool reference](/docs/agents/tools) lists the 21 tools of the default catalogue.