Playwright MCP, Explained by People Who Test for a Living

What is Playwright MCP?

MCP and Playwright MCP are two different things that get mixed up.

• MCP (Model Context Protocol) is an open protocol, created by Anthropic, that lets AI apps call external tools in a standard way.
• Playwright MCP is one MCP server, built by Microsoft's Playwright team and published as @playwright/mcp. It exposes a real browser as a set of tools.

When your agent wants to see a page, Playwright MCP returns an accessibility snapshot: a structured list of what's on the page, with a reference for each element. The model reads that text and picks the next action. That approach needs no vision model, which is why it's fast and fairly reliable on normal web pages.

Should you use Playwright MCP?

In our view, yes, if you want your coding agent to look at and interact with a running app while you work. No, if you want a test suite that gates your releases.

How do you set up Playwright MCP?

You need Node.js 18 or newer. Then add the server to your client.

claude mcp add playwright npx @playwright/mcp@latest

codex mcp add playwright npx "@playwright/mcp@latest"

code --add-mcp '{"name":"playwright","command":"npx","args":["@playwright/mcp@latest"]}'

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    }
  }
}

What can Playwright MCP do?

We grouped the tools by the job you're doing, working from Microsoft's current README.

Tool names change between versions. If a prompt mentions a tool that doesn't exist, run /mcp in Claude Code to see the live list.

Which prompts work best with Playwright MCP?

Vague prompts burn tokens. Name the URL, the steps and what "correct" looks like.

Reproduce a bug

Open http://localhost:3000/invoices. Create an invoice for "Acme Ltd" with two line items, $120 and $80. Apply a 10% discount. Tell me the total shown, and paste any console errors.

Check your own change

I just changed the date picker on /bookings. Open it, pick next Friday, submit, and confirm the confirmation page shows the same date. Take a screenshot of the confirmation.

Find out why something is slow

Load /dashboard and list every network request that takes longer than 1 second, with its URL and status.

Test an error path with mocking (needs --caps network)

Make requests to /api/orders return a 500. Load /orders and tell me exactly what the user sees.

Draft a test, don't trust it

Walk through password reset on staging, then write a Playwright test in TypeScript for that flow using role-based locators. Don't run it, I'll review it.

Where does Playwright MCP fall short?

These limits come from using it, and Microsoft's own README confirms several of them.

1. It's token-heavy. Every snapshot of a busy page goes into the model's context, and context is re-sent on every turn. Microsoft's README says Playwright CLI with skills is more token-efficient for coding agents for this reason. Slack Engineering measured roughly 3.5 to 3.8 million tokens for one complex flow through Playwright MCP, and $15 to $30 per agentic run.

2. The agent grades its own work. When the run ends, the model tells you whether it worked. That's useful while you're developing, but we wouldn't base a release decision on it.

3. Every rerun starts from zero. The agent re-reads and re-reasons through the same flow each time, so the second run is no faster or cheaper than the first.

4. One browser per profile. Only one browser can use a persistent profile at a time, so parallel sessions need --isolated or separate --user-data-dir values.

5. The accessibility tree has blind spots. Canvas, WebGL, maps and some custom components show up poorly or not at all. The vision capability helps, at the cost of coordinate-based clicking.

6. It isn't a security boundary. Microsoft says so plainly in the README.

Is Playwright MCP safe to use?

It's as safe as the browser you give it. A few rules we follow:

• Use --isolated by default. A persistent profile means the agent is signed in to whatever you were signed in to.
• Never let it browse untrusted sites while logged in. Text on a page can contain instructions aimed at the model. This is prompt injection, and the browser can't tell the difference.
• Keep secrets out of prompts. Load sessions with --storage-state from a test account, not your own login.
• Watch browser_evaluate and browser_run_code_unsafe. They run JavaScript in the page. Read what the agent wants to run before you approve it.
• Point it at staging, not production, unless the task is read-only.

How much does Playwright MCP cost to run?

The server is free and open source. You pay for model tokens, and three things drive the bill: how big each snapshot is, how many turns the task takes, and which model you use. To cut it:

• Write specific prompts so the agent takes fewer turns.
• Start on the exact URL, not the homepage.
• Enable only the --caps you need, since every tool's description sits in context.
• Close sessions when you're done. Long sessions keep re-sending old snapshots.

Playwright MCP vs Playwright CLI vs Chrome DevTools MCP vs Kane CLI

These tools get compared as if they compete, but we think they mostly do different jobs.

What are Playwright's test agents?

Playwright also ships three test agents (a planner, a generator and a healer) that you install into your agent with npx playwright init-agents --loop=claude. The planner explores your app and writes a test plan, the generator turns it into Playwright test code, and the healer tries to repair failing tests. They produce code that you own and maintain. We also have a guide to help you understand how to use Playwright MCP for AI Agent workflows step by step, from a Jira ticket to a self-healing test run.

Where does Kane CLI fit with Playwright MCP?

We recommend Playwright MCP to explore and debug, and Kane CLI when you need a verdict.

Before you ship, you need a pass or fail that the agent doesn't get to decide, that runs the same way in CI, and that leaves evidence someone else can check. Playwright MCP leaves that gap, and Kane CLI is built for it. For pass or fail results, we think Kane CLI is the best alternative to Playwright MCP. You describe the flow in plain English, it drives a real Chrome, and it returns a typed result.

npm install -g @testmuai/kane-cli

kane-cli run "Create an invoice for Acme Ltd with items of 120 and 80, \
  apply a 10% discount, and verify the total shows 180.00" \
  --url http://localhost:3000/invoices --agent --headless \
  | tail -1 | jq '{status, one_liner, test_url}'

The agent gets one line back instead of a stack of page snapshots. Exit code 0 is a pass, 1 a failure, 2 a setup error, 3 a timeout.

A workflow that uses both:

• Explore with Playwright MCP. Ask your agent to reproduce the bug and find the cause.
• Fix the code.
• Verify with Kane CLI. Your agent runs kane-cli run "..." --agent and only says "done" on exit code 0.
• Keep it. Save the flow as a _test.md file. Kane CLI replays it on every pull request with kane-cli testrun run --parallel 4 and re-authors only the steps the UI changed.

To let your coding agent call Kane CLI on its own, install the skill once:

npx @testmuai/kane-cli-skill

It works with Claude Code, Codex CLI and Gemini CLI.

Can Playwright MCP run in CI?

It can, but in our opinion it's rarely the right call. A CI step needs a fixed pass/fail rule, a time limit and a cost you can predict. An open-ended agent session gives you none of the three. Use Playwright MCP while developing. In CI, run Playwright test code or a verifier with exit codes, such as Kane CLI.

Install Kane CLI