Playwright MCP vs a hosted browser MCP server: when to use which

Short answer: use Microsoft's Playwright MCP (free, open source) when your coding agent needs to drive a browser on your own machine, for example to test your app on localhost step by step. Use a hosted, task-level MCP server such as Human Browser when the target is a real public site that blocks datacenter IPs or headless browsers, when you need logged-in profiles, a human for an OTP or captcha, or many runs in parallel off your laptop. Your agent sends one goal; a real Chrome on a residential IP does the clicking and returns the result. Both can be installed side by side.

No subscription. Prepaid balance, first top-up can be $1. See pricing.

Two different shapes of browser MCP

Playwright MCP is a local browser with step-level tools. Your coding agent (Claude Code, Cursor, Codex, Cline) calls a tool to open a page, reads back the page state, decides what to click or type, calls the next tool, and so on. You get full control and it costs nothing, but every step is a round trip through your agent, the page state lands in its context window, and the browser runs from your IP with whatever proxies and captcha handling you set up yourself.

Human Browser MCP is a hosted, task-level server. It exposes three tools: humanbrowser_run starts a task from a goal in plain English and returns a task id and a live viewer URL within seconds; humanbrowser_status long-polls (up to 60 seconds per call) until the task is completed, failed or needs a human; humanbrowser_viewer_url returns the viewer link on its own. Our agent drives a real headful Chrome in the cloud on a residential IP, handles captchas, and hands over to a human in the viewer when it hits something like an OTP. Your coding agent only sees the task status and the final answer.

Playwright MCP vs Human Browser MCP

Human Browser cost figures: measured on our production over 132 successful customer tasks in September 2026 (median task about 8agent steps and 4.5 billed minutes, median all-in cost $0.24 at today's rates; p25 $0.12, p75 $0.39). Sessions under 30 seconds and time while a human has taken over are not billed. Residential traffic is $4/GB.

When Playwright MCP is the better choice

• You are testing your own app. A hosted browser cannot reach localhost on your laptop. For writing and debugging end-to-end tests, checking a UI change, or reproducing a bug in your dev server, a local browser is the right tool.
• You want the agent to control every step. With step-level tools your coding agent decides each click and you can see exactly why. That is useful when you are building a Playwright script and want to turn the steps into code.
• The target does not fight automation. Internal tools, staging sites, docs pages and friendly public pages usually work fine from your own IP.
• Cost has to be zero. Playwright MCP is free. Human Browser costs $0.10 per browser-hour plus $0.02 per agent step, traffic and captchas, which is more than running a browser yourself.
• You need raw browser access. Human Browser does not hand you a Playwright or CDP connection; it takes a goal and returns a result.

When a hosted browser MCP is the better choice

• The site blocks datacenter IPs or headless browsers. Human Browser runs a real headful Chrome on a residential exit by default, built to pass the checks that stop a local headless run or a cloud VM.
• Captchas and OTPs show up. About 13% of our customers’ successful tasks met a captcha; they are handled automatically at $0.005 each. When the agent needs a code or a decision, the task moves to input-required and a human can finish the step in the live viewer.
• You need to stay logged in. Pass a named profile and cookies and logins persist across runs.
• You want runs off your laptop, in parallel. Up to 10 concurrent sessions per paid token, and they keep going when you close the lid.
• You want to keep your coding agent's context clean. Page snapshots stay on our side; your agent gets a status and a result instead of the page state from every step.

Set it up in Claude Code, Cursor, Codex or Cline

Get a token first (it looks like hb_live_...), then add the server. Full per-client instructions are in the MCP docs.

Remote server (streamable HTTP)

For clients that accept a URL, such as Cursor and Cline:

{
  "mcpServers": {
    "humanbrowser": {
      "url": "https://agent.humanbrowser.cloud/mcp",
      "headers": { "Authorization": "Bearer hb_live_..." }
    }
  }
}

Claude Code

Add the remote server from the terminal:

claude mcp add --transport http humanbrowser \
  https://agent.humanbrowser.cloud/mcp \
  --header "Authorization: Bearer hb_live_..."

Local variant (npx, stdio)

For clients that start MCP servers as a local command:

{
  "mcpServers": {
    "humanbrowser": {
      "command": "npx",
      "args": ["-y", "@virixlabs/humanbrowser", "mcp"],
      "env": { "HB_TOKEN": "hb_live_..." }
    }
  }
}

Example prompt to your coding agent

Use the humanbrowser MCP tool. Start one task with humanbrowser_run:
"Open the pricing pages of these three competitors, and for each plan
return name, monthly price and included seats as JSON."
Use country "us". Send me the viewer_url so I can watch.
Poll humanbrowser_status with wait_seconds 30 until the state is
completed, failed or input-required. If input-required, tell me
what it needs. When completed, save the JSON to data/competitors.json.

Your agent calls humanbrowser_run once, passes you the viewer link, and polls humanbrowser_status until the result is ready. If you also have Playwright MCP installed, you can tell the agent which one to use for which part of the job: for example Playwright MCP to check your local build, Human Browser for the competitor sites.

FAQ

Related