The Best MCP Servers to Control a Browser From Your Coding Agent (2026)
By Steven Van ·
Browser Use compared five browser-control tools for coding agents; four were tested on 100 tasks, and scores were statistically tied while token use differed.
Browser Use has published a comparison of five ways to let a coding agent such as Claude Code, Codex or Cursor drive a browser, and tested four of them on a matched eval. The company makes Browser Use and Browser Harness, and says so in the post, which puts Browser Harness first. Competitor facts come from their own repos and docs, checked October 18, 2026.
The five options
- Browser Harness: runs in the Chrome you already have open, logged in as you. It talks to Chrome over the Chrome DevTools Protocol, with no Playwright and no extension.
- Playwright MCP: for clean throwaway sessions, and for Firefox and WebKit.
- Chrome DevTools MCP: for debugging your own app's performance, network or console.
- agent-browser: one binary with an encrypted auth vault, a domain allowlist and action policies.
- Browser MCP: the quickest way to hand one tab to an agent, but its MCP server has not had an npm release since April 2025. It was not part of the eval.
How the eval was run
Four tools ran on 100 public tasks from BU Bench V2.1 (bu2-001 to bu2-100, half of its 200). Every tool used the same agent, Codex CLI 0.144.6 with GPT-6 Luna on medium reasoning with web search off, and the same fresh Browser Use cloud browser with a US proxy. The prompt was identical except for one line naming the tool, and the same frozen V2.1 judge scored every run. Agent failures were never retried. Runner, browser and judge failures were rerun once. It was measured October 14, 2026.
Playwright MCP and Chrome DevTools MCP ran as MCP servers with default settings. Browser Harness ran as its CLI and skill, not as browser-harness-mcp, and agent-browser also ran as its CLI and skill.
Results
All figures are per task. Agent cost is the model bill only, not browser, proxy or judge.
- Browser Harness: score 60.4, 11% fully solved, 2.1M median input tokens, 42 median tool calls, $0.042, 9.0 min.
- Playwright MCP: score 60.0, 11%, 4.4M tokens, 57 calls, $0.066, 8.3 min.
- Chrome DevTools MCP: score 61.0, 11%, 6.7M tokens, 67.5 calls, $0.097, 8.7 min.
- agent-browser: score 56.3, 9%, 3.3M tokens, 45 calls, $0.056, 11.3 min.
Browser Use says nobody won on score. Paired differences, Browser Harness minus each tool, were +0.4 points against Playwright MCP, -0.6 against Chrome DevTools MCP and +4.0 against agent-browser. Every 95% interval includes zero. Browser Harness used the fewest tokens and was cheapest per task. Playwright MCP and Chrome DevTools MCP were faster, and agent-browser was slowest.
The post does not know why the token gap exists, since it was not isolated. It suggests MCP overhead may account for part of it, but notes agent-browser is also a CLI and still used about 1.6x Browser Harness's tokens. browser-harness-mcp was not run, so there is no token count for it.
Limits of the test
The eval used one run per task, 100 of 200 tasks, one agent and one model, and a fresh cloud browser every time rather than a logged-in Chrome. A first agent-browser pass was discarded because a setup bug let it skip the shared cloud browser, and the table shows the corrected rerun.
Browser Harness setup and caveats
The measured setup is the CLI plus a skill, which an agent can install from a pasted prompt. Chrome needs a one-time step: tick "Allow remote debugging for this browser instance" at chrome://inspect/#remote-debugging. On Chrome 144+ you then click Allow once per connection, and on macOS browser-harness mac-approve clicks it for you. For MCP-only clients such as Cursor, browser-harness-mcp exposes 23 tools, though that path was not measured.
It is MIT licensed and free in your own Chrome with no API key. The CLI can also start Browser Use cloud browsers at $0.02 per browser-hour, with residential proxy traffic at $5/GB on by default.
Listed caveats: the agent sees the whole logged-in Chrome, one daemon has one attached tab so concurrent agents can race, and no clean-profile mode is documented for the MCP server. There is no Firefox or WebKit support. PyPI is at 0.1.13 from September 4, behind the main branch, the package is marked alpha, and telemetry is on by default (browser-harness telemetry disable turns it off).