Agentic WeeklyGet the weekly

Playwright MCP server: setup, tools and what we measured

Playwright MCP is Microsoft's MCP server that lets an AI agent drive a real browser. In our 2026-10-25 smoke test, version 0.0.83 started in 383 ms, listed 25 tools and navigated to a local page in 40.5 ms median, with no language model involved. We did not test clicking, typing or screenshots.

By Agentic Weekly, operated independently from Texas, US · Last updated (America/Chicago) · How we work

Checked by Agentic Weekly's editorial process: vendor docs re-read (the commands were run on 2026-10-25, not re-run on this date).

Last verified: 2026-10-26 (America/Chicago) · Versions covered: @playwright/mcp 0.0.83 (the server prints Playwright 1.64.0-alpha-1790635538000); Node 20.19.2 and 24.22.1 · Evidence: Tested (see “What we tested”)

What Playwright MCP is and who maintains it

Playwright MCP is an MCP server from Microsoft that exposes a browser to an AI agent. The README (read 2026-10-26) says it uses Playwright's accessibility tree rather than pixel input, and that it needs no vision model. The npm package @playwright/mcp lists the Apache-2.0 license and node >=18 (npm registry, read 2026-10-26). The latest GitHub release when we read it was v0.0.83, published 2026-09-28.

New to the term? What is an MCP server? explains hosts, clients and servers. The README also states that Playwright MCP is not a security boundary, so treat any page the agent opens as untrusted input.

Prerequisites

  • Operating system and hardware we ran on: Linux x86_64 virtual machine (Debian 13), 8 vCPU, 15 GiB RAM, no GPU. Not run by us: macOS and Windows.
  • Node.js: we ran 20.19.2 and 24.22.1 (both worked). The README says Node.js 18 or newer; we did not run Node 18.
  • A browser: the package does not bundle one. We used the machine's existing Google Chrome at /usr/bin/google-chrome.
  • Accounts and keys: none. Network: npm registry access to install.
  • A model is not needed to start the server. Our measurements used a small Python client written with the official MCP Python SDK 2.3.0, with no language model involved.

Step 1: Install the package

Install @playwright/mcp with npm. Our run resolved version 0.0.83; the latest version may differ when you read this.

Command (as logged)
npm install --no-audit --no-fund --cache /tmp/npmcache @playwright/mcp
Expected result (from the log notes)
Resolved 0.0.83; 19 MB in 3 packages.
Install time: 1.3 s on Node 20.19.2, 1.0 s on Node 24.22.1 (cold npm cache).

The small size is because no browser is downloaded. The server binary is node_modules/.bin/playwright-mcp.

Step evidence: Tested (npm install on both Node versions, 2026-10-25; log: /tests/playwright-mcp).

Step 2: Start the server in headless Chrome

Point the server at an installed Chrome and run it without a window. These are the exact flags in our log; each is explained in the flags table below.

Command (as logged)
node node_modules/.bin/playwright-mcp --headless --isolated --browser chrome --executable-path /usr/bin/google-chrome --no-sandbox
Expected result (what our client saw during the handshake)
server: Playwright 1.64.0-alpha-1790635538000
protocol: 2025-11-25
cold start (spawn to initialize result): 383 ms

The server speaks MCP over stdio, so you normally let your MCP client launch it rather than running it by hand. Remove --no-sandbox unless you are root or in a container (see Troubleshooting).

Step evidence: Tested (start over stdio with these flags, 2026-10-25; log: /tests/playwright-mcp).

Step 3: Connect it to an MCP client

These are the setup lines the README gives for popular clients (read 2026-10-26). We did not run any of them; our own test launched the server from a script.

Setup by client, from the Playwright MCP README
ClientCommand or configSource and dateRun by us?
Claude Codeclaude mcp add playwright npx @playwright/mcp@latestREADME, read 2026-10-26No
VS Codecode --add-mcp '{"name":"playwright","command":"npx","args":["@playwright/mcp@latest"]}'README, read 2026-10-26No
CursorCursor Settings, MCP, Add new MCP Server, type command, command npx @playwright/mcp@latestREADME, read 2026-10-26No
Claude Desktop and most othersThe README's "standard config": an mcpServers entry named playwright with "command": "npx" and "args": ["@playwright/mcp@latest"]README, read 2026-10-26No

To pass flags from Claude Code, the Claude Code docs say everything after -- goes to the server untouched, so the shape is claude mcp add playwright -- npx @playwright/mcp@latest --headless. That line combines the two documents and we have not run it. For other Claude Code options see Add an MCP server to Claude Code.

Step evidence: Reported (not run by us). Source: https://github.com/microsoft/playwright-mcp and https://code.claude.com/docs/en/mcp, read 2026-10-26.

Step 4: See what the server exposes

In our run tools/list returned 25 tools in 10.1 ms. The README also documents extra tool groups that are switched on with --caps (for example vision, pdf, devtools and testing); we did not enable any, so 25 is the default set with the flags above and can change by version.

Tool names printed by tools/list in the same run
browser_close, browser_resize, browser_console_messages, browser_handle_dialog, browser_emulate_media, browser_evaluate, browser_file_upload, browser_drop, browser_find, browser_fill_form, browser_press_key, browser_type, browser_navigate, browser_navigate_back, browser_network_requests, browser_network_request, browser_run_code_unsafe, browser_take_screenshot, browser_snapshot, browser_click, browser_drag, browser_hover, browser_select_option, browser_tabs, browser_wait_for

The tool count matters because every tool schema takes space in a model's context window. We did not measure that cost.

Step evidence: Tested (tool count and timing, 2026-10-25; log: /tests/playwright-mcp).

Step 5: Make one call and read the result

We timed browser_navigate to a local fixture page served on 127.0.0.1. Ten sequential calls gave a median of 40.5 ms and a 95th percentile of 51.8 ms.

Output of one timed call (first 300 characters, from the log)
browser_navigate({"url": "http://127.0.0.1:8099/test.html"})
### Ran Playwright code
```js
await page.goto('http://127.0.0.1:8099/test.html');
```
### Page
- Page URL: http://127.0.0.1:8099/test.html
- Page Title: Fixture page
### Snapshot
- [Snapshot](.playwright-mcp/page-2026-10-26T03-39-23-508Z.yml)

The result carries the page title and a reference to a saved accessibility snapshot. That is a protocol and process timing for a tiny local page on a shared machine, not a measure of real-world page loads.

Step evidence: Tested (10 calls to browser_navigate, 2026-10-25; log: /tests/playwright-mcp).

Flags that matter

Playwright MCP flags: what the README says and whether our run used them
FlagWhat the README saysUsed in our run
--headlessRun the browser in headless mode; the default is headed.Yes
--isolatedKeep the browser profile in memory and do not save it to disk. The README says concurrent clients sharing a persistent profile conflict, and suggests --isolated or a distinct --user-data-dir for parallel clients.Yes
--browser <browser>Browser or Chrome channel to use; possible values are chrome, firefox, webkit and msedge.Yes (chrome)
--executable-path <path>Path to the browser executable.Yes (/usr/bin/google-chrome)
--no-sandboxDisable the sandbox for all process types that are normally sandboxed.Yes (we ran as root on a VM)

Flag descriptions are quoted in substance from the README, read 2026-10-26. We only list flags that appear in the README and in our log; the README documents many more.

What we tested

Evidence: Tested (smoke test, 2026-10-25, no LLM involved) for the steps marked Tested; Reported (not run by us) for client setup.

  • What we ran: the install and start commands above, then a small Python client (official MCP Python SDK 2.3.0) over stdio that timed the handshake, one tools/list and 10 sequential browser_navigate calls. Server string printed: Playwright 1.64.0-alpha-1790635538000 (npm package 0.0.83). Run on 2026-10-25 at 22:39 CDT (America/Chicago).
  • Environment: Linux x86_64 virtual machine (Debian 13), 8 vCPU, 15 GiB RAM, no GPU, shared with other jobs.
  • Model: no language model involved.
  • Full test log: Playwright MCP test log.
  • Reported, from the README and Claude Code docs, not run by us: the client setup lines in Step 3 and the flag descriptions.
  • Not tested: remote (HTTP) transport, authentication and concurrent clients; large payloads; clicking, typing, screenshots and snapshot-based element selection (only navigation was timed); any model driving the server; answer quality; Node 18.
Evidence by step
StepEvidenceCovered by
1 InstallTested/tests/playwright-mcp
2 Start with flagsTested/tests/playwright-mcp
3 Connect a clientReportedREADME, read 2026-10-26
4 List toolsTested/tests/playwright-mcp
5 One navigate callTested/tests/playwright-mcp

This is a smoke test, not a benchmark: one client, one tiny local page, one shared machine. A single run per Node version was used for install; the latency run used 10 calls.

Troubleshooting

Playwright MCP problems and fixes
SymptomLikely causeFixSource
Chrome will not start as root or in a containerChrome's sandbox needs privileges that root or a container lacksPass --no-sandbox. Do not do this on a machine that browses untrusted pages.Gotcha noted in our log 2026-10-25; we passed the flag and did not record the failure without it, /tests/playwright-mcp
The server reports Playwright 1.64.0-alpha-... but you installed 0.0.83The package version and the Playwright build it embeds are different numbersNothing to fix. Quote both when you report a problem.Seen by us 2026-10-25, /tests/playwright-mcp
Two clients fight over one browser profileA persistent profile can be used by only one browser instance at a timeStart each extra client with --isolated or a distinct --user-data-dir.README, read 2026-10-26
The client cannot find the serverThe client was configured with a command that is not on its PATHUse the full path to npx or to node_modules/.bin/playwright-mcp in the client config.Not reproduced by us; general MCP client behaviour, check your client's docs

The last row has no vendor source, so treat it as a hint only.

Playwright MCP vs Chrome DevTools MCP

Both servers drive Chrome, but they expose different tools and we timed different calls. The comparison page covers the rest: Playwright MCP vs Chrome DevTools MCP. For the agent-side alternative see Browser Use vs Playwright MCP.

Playwright MCP and Chrome DevTools MCP in our 2026-10-25 runs
AttributePlaywright MCPChrome DevTools MCP
Tools listed2532
Call we timedbrowser_navigate, 10 calls, median 40.5 msnew_page, 10 calls, median 136.6 ms
Evidence labelTested (smoke test, 2026-10-25)Tested (smoke test, 2026-10-25)

The two timings are for different calls and are not comparable. Logs: Playwright MCP and Chrome DevTools MCP.

  • Playwright MCP: Microsoft's MCP server that lets agents drive a browser with Playwright. Tested (smoke test, 2026-10-25, no language model involved).

FAQ

How do I install the Playwright MCP server?

Run npm install @playwright/mcp. On 2026-10-25 that resolved version 0.0.83 (19 MB, 3 packages) in about a second on Node 20.19.2 and 24.22.1. A browser is not bundled, so we pointed it at an installed Google Chrome with --executable-path.

How many tools does Playwright MCP have?

Our tools/list call on 2026-10-25 returned 25 tools for version 0.0.83 with the flags in this guide. The README documents extra groups behind the --caps flag, and the count can change by version.

Does Playwright MCP work with a local model?

We timed the server without a language model; we have not yet measured a local model driving it. The server exposes 25 tool schemas, which is a lot of context for a small model, but we did not test that.

Is Playwright MCP free?

The npm package lists the Apache-2.0 license (npm registry, read 2026-10-26), so the server itself is open source. Costs can come from the model or services you connect it to; we did not check any pricing.

Do I need --no-sandbox?

We passed it because we ran as root on a virtual machine, and our log notes Chrome needs it there. It disables Chrome's sandbox, so do not use it on a machine that opens untrusted pages. As a normal user you may not need it; we did not test that.

Have you tested Playwright MCP?

Yes, partly. On 2026-10-25 we ran a no-model smoke test: install, start, tools/list (25 tools) and 10 browser_navigate calls. We did not test clicking, typing, screenshots, remote transport or any model driving it. Client setup is from the README.

Sources

Page history

  • : First published. README, npm registry and release page read on this date; commands come from the 2026-10-25 test log and were not re-run.

Markdown version for agents: /guides/playwright-mcp-server.md