The site is built by a machine of agents

mcp server · mcp catalog · snapshot Jul 2

Playwright MCP

Q: What is Playwright MCP in simple terms?

It is an AI agent's “hands and eyes” in the browser: an official Microsoft MCP server through which Claude Code, Cursor, or Codex open pages, click, and read content. Built on the Playwright engine for e2e testing.

Q: How is it better than screenshot-based browser control?

Playwright MCP reads the accessibility tree - a structured page description. The agent does not need to “look at” pixels: elements are addressed precisely, it works faster and more reliably, and no vision model is needed. Screenshot mode is available as an opt-in (--caps=vision).

Q: How many tokens does it consume, and how can I reduce it?

Tool definitions ≈ 13.7K tokens; a 10-step task ≈ 114K through MCP versus 27K through Playwright CLI (Microsoft benchmark). Reduce usage with minimal --caps, targeted snapshots, and writing tests through CLI.

Q: How does Playwright MCP differ from Playwright itself?

Playwright is an automation library controlled by your code. Playwright MCP is a server built on top of it, controlled by an agent via the MCP protocol: the same browsers, but the “steering wheel” is held by an LLM rather than a script.

Q: Playwright MCP or Chrome DevTools MCP?

Playwright - browser driving (scenarios, forms, e2e); Chrome DevTools MCP - diagnostics (perf traces, network, console). Steve Kinney's formula: “driving vs debugging”. Both are often installed.

Q: Is it safe?

The server is local, requires no keys, and has open code. But Microsoft itself calls it “not a security boundary”: the agent controls a real browser with your sessions. The main risk is prompt injection from untrusted pages. Isolate the profile (--isolated) and do not mix “browsing the internet” with access to secrets.

Q: Can I connect it to my already running Chrome?

Yes: the --extension flag connects the server to a live Chrome/Edge through the Playwright Extension, while the --cdp-endpoint flag connects to any browser via CDP.

Q: Does it work from Russia?

Yes, completely: it is a local npm package, with no keys or payment required, and it does not send data externally. The only limitation comes from the client itself (Claude/Cursor), not the server.

Q: Headless or with a window?

By default - with a window (you can see what the agent does). In CI/Docker, enable --headless; the Docker image supports headless chromium only.

Microsoft · Playwright team; namespace verified through GitHub by the registry

68 toolsnpm · stdioheadless / CIchrome · firefox · webkit

Gives an AI agent control of a real browser: open a page, click, fill out a form, read the console and network. Works through the accessibility tree - the page's structural tree, without pixel recognition: fast, deterministic, no vision model needed.

installation · Claude Code

$ claude mcp add playwright npx @playwright/mcp@latest

available from Russia · no keys or card

89.8

score v0 · #16 of 14522

34,583

GitHub stars · owner verified

0.0.77

Registry version

tools · 23 core + 45 opt-in

numbers - official mcp registry + github · snapshot Jul 2

MCP server race · npm downloads per day · live

Playwright MCP7.6M/week

Chrome DevTools1.44M/week

Context7496K/week

1.0

Install or not

Install for

+check layouts hand-coded by the agent: clicks, forms, navigation
+e2e scenarios and bug reproduction in a clean profile
+frontend debugging: console, network, screenshot - from chat
+CI and headless: the only one among agent browser tools

Do not install for

−production scraping - use Firecrawl/crawlers
−sites with anti-bot protection: it does not hide itself (stealth forks exist)
−a one-off check in your logged-in Chrome - native tools are simpler

Free. No API keys needed: the server is local (an npm package), the browser is yours.

2.0

Installation by client

Codex

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

or in ~/.codex/config.toml: [mcp_servers.playwright] command = "npx", args = ["@playwright/mcp@latest"]

Cursor

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

Settings, MCP, Add new global MCP server; or the deeplink button in the project README

VS Code

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

Docker (headless chromium)

{
  "mcpServers": {
    "playwright": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "--init", "--pull=always", "mcr.microsoft.com/playwright/mcp"]
    }
  }
}

No keys or env variables needed. Just one requirement: Node.js 18+. On first launch, Playwright may ask to install a browser: npx playwright install chrome.

useful flags

`--browser <b>`	chrome \| firefox \| webkit \| msedge
`--headless`	without a browser window (with a window by default)
`--caps <c>`	enable opt-in packages: vision, pdf, devtools
`--isolated`	in-memory profile, do not save to disk
`--device <d>`	device emulation, for example "iPhone 15"
`--allowed-origins`	limit where the browser can go (not a security boundary!)
`--extension`	connect to an ALREADY running Chrome/Edge through an extension

availability in Russia

Works from Russia without restrictions: the server is local and sends no data out (the browser downloads from the Playwright CDN - available). No foreign card needed.

3.0

Tools and context cost

68 tools: 23 in the core by default + 45 in optional packages, to avoid bloating context.

Context cost

Tool definitions ≈ 13.7K tokens (ytyng benchmark, 03.2026); a 10-step task ≈ 114K through MCP versus 27K through Playwright CLI (Microsoft measurement); a full-page screenshot - up to 200K+. Reduce: minimum --caps, targeted snapshots, tests - through CLI. Check yours: /context.

▸Core (default)23 tools

browser_navigate / _navigate_backopen URL, go back
browser_snapshotpage accessibility snapshot - the agent's main "eye"read-only
browser_click / _type / _hover / _drag / _dropclicks, text input, hover, drag-n-drop
browser_fill_form / _select_option / _file_uploadforms, selects, file uploads
browser_take_screenshotpage or element screenshotread-only
browser_console_messagesconsole messages (by level)read-only
browser_network_request / _network_requestsrun a request; list page requests
browser_evaluate / _run_code_unsaferun JS on the page (careful: unsafe)
browser_tabstabs: open/close/switch
browser_wait_for / _press_key / _resize / _handle_dialog / _closewaits, keys, window size, dialogs

▸Storage: cookies, localStorage, sessionStorage17 tools · --caps=storage

▸DevTools: tracing, video, element highlighting11 tools · --caps=devtools

▸Vision: coordinate clicks (for screenshot-based models)6 tools · --caps=vision

▸PDF: generate a PDF from a page6 tools · --caps=pdf

▸Network: intercept and mock requests (route)4 tools · --caps=network

▸Config: read configuration on the fly1 tools · --caps=config

4.0

Why use it if Claude Code and Codex have their own browsers?

Not "instead of", but a different task class: native tools cover "check in my logged-in Chrome without setup", while Playwright MCP is the only one with headless/CI and Firefox/WebKit. Facts as of 07.02.2026.

criterion	Playwright MCP	Native (Claude in Chrome / Codex)
Headless / CI / server without GUI	yes: --headless, Docker image, standalone HTTP	no: Claude in Chrome and the Codex browser require a live window
Your logged-in profile	not out of the box; possible through --extension or --cdp-endpoint	Claude in Chrome - yes, out of the box; the Codex in-app browser - without auth/cookies at all (for logins - its Chrome extension)
Browsers	chrome, edge, firefox, webkit	Chrome/Edge only
Clean room (profile isolation)	yes: --isolated, reproducible runs	no - always your live profile
Device emulation, network mocks	yes: --device "iPhone 15", route interception	no
Access price	free, open source	Claude in Chrome - paid plans only (Pro/Max/Team), does not work through Bedrock/Vertex

→Choose Playwright MCP: CI and headless runs, reproducible e2e, Firefox/WebKit, a clean room for bug reproduction, device emulation - or if your agent has no browser of its own (Codex CLI, Gemini CLI).
→Choose Claude in Chrome: the task lives in your logged-in Chrome (Gmail, internal tools), zero setup - /chrome in Claude Code.
→Choose Chrome DevTools MCP: diagnostics, not driving - performance traces, network, console ("debugging, not driving").
→For writing tests at scale - Playwright CLI is about 4 times cheaper in tokens (Microsoft measurement); keep MCP for interactive scenarios.

5.0

Security and trust

Publisher

Microsoft - Playwright team; namespace verified through GitHub by the registry. Stars counted after repository ownership verification.

Authentication

No authentication: a local stdio process launched by the client itself. No keys - nothing to leak.

Verification status

vibecoding.tech security scan: not conducted (honest v1 status). Publisher - Microsoft, code is open (Apache-2.0), 34.5K stars, releases almost daily.

what the server can do

—controls the browser on your behalf: it can open any site, click, submit a form - including where you are logged in (the default profile is saved to disk; --isolated keeps it in memory)
—reads files only within workspace roots; navigation to file:// is blocked by default (can be lifted with --allow-unrestricted-file-access - do not do this without a reason)
—can run arbitrary JS on the page (browser_evaluate, browser_run_code_unsafe)
—there are no outbound "phone home" calls in the code: only the browser itself goes out, on agent commands

candid caveats

- Microsoft explicitly writes: "Playwright MCP is not a security boundary". The --allowed-origins/--blocked-origins flags are convenience filters; they do not apply to redirects
- the main real risk - prompt injection from pages: an untrusted site can "command" the agent in text. Do not let an agent with this server access untrusted sites in a session that has access to secrets
- in CI/Docker, use headless + an isolated profile; --no-sandbox - only inside a container
- known incident: DNS rebinding (issue #1206), closed in v0.0.40 - the server is actively maintained and patched, but keep an eye on the version

what practitioners say · no links to x (blocked in Russia)

"Authentication is easy: have it show you a login page, then log in yourself". Caveat: explicitly say "playwright mcp" in the first prompt - otherwise the agent will use Bash.

Simon Willison @simonw · July 2025 · via til.simonwillison.net

Part of "agentic manual testing": the agent must check its own code rather than assume it works.

Simon Willison @simonw · March 2026 · via simonwillison.net

"Almost all MCPs really should be CLIs" - dropped Playwright MCP in favor of chrome-devtools-mcp: context tax, the agent "ran Playwright where it could simply read the code".

Peter Steinberger @steipete · October 2025 · via steipete.me

Instead of 25+ tools, created playwrightess: one playwright_eval tool that runs JS in a persistent browser session.

Armin Ronacher @mitsuhiko · October 2025 · via lucumr.pocoo.org

"Constantly blows up the context window on nearly every call"; Microsoft confirmed this and recommends a CLI workflow for writing tests (issues #889, #1040).

Ask HN news.ycombinator.com · October 2025 · via news.ycombinator.com/item?id=45764043

6.0

Common issues and fixes

01“Executable doesn't exist” error / browser not found

Run npx playwright install chrome (or chromium). The Docker version downloads the browser itself.

02Client cannot see the server (0 tools)

Check in order: is the config JSON valid; is it the right config file (global vs project); fully restart the client. 90% of cases are one of these three.

03Windows: server does not start through npx

Wrap the command: "command": "cmd", "args": ["/c", "npx", "@playwright/mcp@latest"]. A classic Windows npx pitfall.

04Browser “busy”: profile is already in use

Close the running Chrome using the same profile or add --isolated (in-memory profile). Or connect to a live browser through --extension.

05Agent “goes blind”: page exists, but snapshot is empty/sparse

Wait for loading (browser_wait_for), check whether the content is in an iframe; for canvas/widget pages, enable --caps=vision and work from the screenshot.

06Consumes context: each step - tens of thousands of tokens

Request targeted element snapshots instead of the full page, do not enable unnecessary --caps, use screenshots only when visual inspection is needed. For writing tests at scale, use Playwright CLI (≈4 times cheaper in tokens, Microsoft measurement).

07Node < 18 or npx pulls an old version

Node 18+ is required. With NVM, specify the absolute path to node in command; @latest in args keeps the server current.

7.0

Frequently asked questions

▸What is Playwright MCP in simple terms?

It is an AI agent's “hands and eyes” in the browser: an official Microsoft MCP server through which Claude Code, Cursor, or Codex open pages, click, and read content. Built on the Playwright engine for e2e testing.

▸How is it better than screenshot-based browser control?

Playwright MCP reads the accessibility tree - a structured page description. The agent does not need to “look at” pixels: elements are addressed precisely, it works faster and more reliably, and no vision model is needed. Screenshot mode is available as an opt-in (--caps=vision).

▸How many tokens does it consume, and how can I reduce it?

Tool definitions ≈ 13.7K tokens; a 10-step task ≈ 114K through MCP versus 27K through Playwright CLI (Microsoft benchmark). Reduce usage with minimal --caps, targeted snapshots, and writing tests through CLI.

▸How does Playwright MCP differ from Playwright itself?

Playwright is an automation library controlled by your code. Playwright MCP is a server built on top of it, controlled by an agent via the MCP protocol: the same browsers, but the “steering wheel” is held by an LLM rather than a script.

▸Playwright MCP or Chrome DevTools MCP?

Playwright - browser driving (scenarios, forms, e2e); Chrome DevTools MCP - diagnostics (perf traces, network, console). Steve Kinney's formula: “driving vs debugging”. Both are often installed.

▸Is it safe?

The server is local, requires no keys, and has open code. But Microsoft itself calls it “not a security boundary”: the agent controls a real browser with your sessions. The main risk is prompt injection from untrusted pages. Isolate the profile (--isolated) and do not mix “browsing the internet” with access to secrets.

▸Can I connect it to my already running Chrome?

Yes: the --extension flag connects the server to a live Chrome/Edge through the Playwright Extension, while the --cdp-endpoint flag connects to any browser via CDP.

▸Does it work from Russia?

Yes, completely: it is a local npm package, with no keys or payment required, and it does not send data externally. The only limitation comes from the client itself (Claude/Cursor), not the server.

▸Headless or with a window?

By default - with a window (you can see what the agent does). In CI/Docker, enable --headless; the Docker image supports headless chromium only.

alternatives

Chrome DevTools MCP →

official Chrome team server: stronger for debugging (profiling, DevTools protocol), weaker as “hands” for scenarios

ByteDance Browser MCP →

alternative agent browser implementation from the ByteDance monorepo

Browserbase / stealth forks

cloud browser and anti-bot bypasses - when scale or concealment is needed; this is no longer a “developer browser”

reg · io.github.microsoft/playwright-mcp

distribution · npm

card rev · 2026-07-02

data snapshot · July 2, 2026

Sources (verified 2026-07-02): README microsoft/playwright-mcp (tools, flags, security) · Official MCP Registry: io.github.microsoft/playwright-mcp entry · npm @playwright/mcp - 5 000 572 downloads per week (22–28.06.2026)

Data: official MCP registry + GitHub API, daily snapshot. Score v0 - methodology · registry entry

Sections

1.0Install or not 2.0Installation by client 3.0Tools and context cost 4.0Why use it if Claude Code and Codex have their own browsers?5.0Security and trust 6.0Common issues and fixes 7.0Frequently asked questions

14,522 servers

mcp catalog

mcp server · mcp catalog · snapshot Jul 2

Playwright MCP

Microsoft · Playwright team; namespace verified through GitHub by the registry

68 toolsnpm · stdioheadless / CIchrome · firefox · webkit

installation · Claude Code

$ claude mcp add playwright npx @playwright/mcp@latest

available from Russia · no keys or card

89.8

score v0 · #16 of 14522

34,583

GitHub stars · owner verified

0.0.77

Registry version

tools · 23 core + 45 opt-in

numbers - official mcp registry + github · snapshot Jul 2

MCP server race · npm downloads per day · live

Playwright MCP7.6M/week

Chrome DevTools1.44M/week

Context7496K/week

1.0

Install or not

Install for

+check layouts hand-coded by the agent: clicks, forms, navigation
+e2e scenarios and bug reproduction in a clean profile
+frontend debugging: console, network, screenshot - from chat
+CI and headless: the only one among agent browser tools

Do not install for

−production scraping - use Firecrawl/crawlers
−sites with anti-bot protection: it does not hide itself (stealth forks exist)
−a one-off check in your logged-in Chrome - native tools are simpler

Free. No API keys needed: the server is local (an npm package), the browser is yours.

2.0

Installation by client

Codex

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

or in ~/.codex/config.toml: [mcp_servers.playwright] command = "npx", args = ["@playwright/mcp@latest"]

Cursor

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

Settings, MCP, Add new global MCP server; or the deeplink button in the project README

VS Code

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

Docker (headless chromium)

{
  "mcpServers": {
    "playwright": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "--init", "--pull=always", "mcr.microsoft.com/playwright/mcp"]
    }
  }
}

No keys or env variables needed. Just one requirement: Node.js 18+. On first launch, Playwright may ask to install a browser: npx playwright install chrome.

useful flags

`--browser <b>`	chrome \| firefox \| webkit \| msedge
`--headless`	without a browser window (with a window by default)
`--caps <c>`	enable opt-in packages: vision, pdf, devtools
`--isolated`	in-memory profile, do not save to disk
`--device <d>`	device emulation, for example "iPhone 15"
`--allowed-origins`	limit where the browser can go (not a security boundary!)
`--extension`	connect to an ALREADY running Chrome/Edge through an extension

availability in Russia

Works from Russia without restrictions: the server is local and sends no data out (the browser downloads from the Playwright CDN - available). No foreign card needed.

3.0

Tools and context cost

68 tools: 23 in the core by default + 45 in optional packages, to avoid bloating context.

Context cost

▸Core (default)23 tools

browser_navigate / _navigate_backopen URL, go back
browser_snapshotpage accessibility snapshot - the agent's main "eye"read-only
browser_click / _type / _hover / _drag / _dropclicks, text input, hover, drag-n-drop
browser_fill_form / _select_option / _file_uploadforms, selects, file uploads
browser_take_screenshotpage or element screenshotread-only
browser_console_messagesconsole messages (by level)read-only
browser_network_request / _network_requestsrun a request; list page requests
browser_evaluate / _run_code_unsaferun JS on the page (careful: unsafe)
browser_tabstabs: open/close/switch
browser_wait_for / _press_key / _resize / _handle_dialog / _closewaits, keys, window size, dialogs

▸Storage: cookies, localStorage, sessionStorage17 tools · --caps=storage

▸DevTools: tracing, video, element highlighting11 tools · --caps=devtools

▸Vision: coordinate clicks (for screenshot-based models)6 tools · --caps=vision

▸PDF: generate a PDF from a page6 tools · --caps=pdf

▸Network: intercept and mock requests (route)4 tools · --caps=network

▸Config: read configuration on the fly1 tools · --caps=config

4.0

Why use it if Claude Code and Codex have their own browsers?

criterion	Playwright MCP	Native (Claude in Chrome / Codex)
Headless / CI / server without GUI	yes: --headless, Docker image, standalone HTTP	no: Claude in Chrome and the Codex browser require a live window
Your logged-in profile	not out of the box; possible through --extension or --cdp-endpoint	Claude in Chrome - yes, out of the box; the Codex in-app browser - without auth/cookies at all (for logins - its Chrome extension)
Browsers	chrome, edge, firefox, webkit	Chrome/Edge only
Clean room (profile isolation)	yes: --isolated, reproducible runs	no - always your live profile
Device emulation, network mocks	yes: --device "iPhone 15", route interception	no
Access price	free, open source	Claude in Chrome - paid plans only (Pro/Max/Team), does not work through Bedrock/Vertex

→Choose Playwright MCP: CI and headless runs, reproducible e2e, Firefox/WebKit, a clean room for bug reproduction, device emulation - or if your agent has no browser of its own (Codex CLI, Gemini CLI).
→Choose Claude in Chrome: the task lives in your logged-in Chrome (Gmail, internal tools), zero setup - /chrome in Claude Code.
→Choose Chrome DevTools MCP: diagnostics, not driving - performance traces, network, console ("debugging, not driving").
→For writing tests at scale - Playwright CLI is about 4 times cheaper in tokens (Microsoft measurement); keep MCP for interactive scenarios.

5.0

Security and trust

Publisher

Microsoft - Playwright team; namespace verified through GitHub by the registry. Stars counted after repository ownership verification.

Authentication

No authentication: a local stdio process launched by the client itself. No keys - nothing to leak.

Verification status

vibecoding.tech security scan: not conducted (honest v1 status). Publisher - Microsoft, code is open (Apache-2.0), 34.5K stars, releases almost daily.

what the server can do

—controls the browser on your behalf: it can open any site, click, submit a form - including where you are logged in (the default profile is saved to disk; --isolated keeps it in memory)
—reads files only within workspace roots; navigation to file:// is blocked by default (can be lifted with --allow-unrestricted-file-access - do not do this without a reason)
—can run arbitrary JS on the page (browser_evaluate, browser_run_code_unsafe)
—there are no outbound "phone home" calls in the code: only the browser itself goes out, on agent commands

candid caveats

- Microsoft explicitly writes: "Playwright MCP is not a security boundary". The --allowed-origins/--blocked-origins flags are convenience filters; they do not apply to redirects
- the main real risk - prompt injection from pages: an untrusted site can "command" the agent in text. Do not let an agent with this server access untrusted sites in a session that has access to secrets
- in CI/Docker, use headless + an isolated profile; --no-sandbox - only inside a container
- known incident: DNS rebinding (issue #1206), closed in v0.0.40 - the server is actively maintained and patched, but keep an eye on the version

what practitioners say · no links to x (blocked in Russia)

"Authentication is easy: have it show you a login page, then log in yourself". Caveat: explicitly say "playwright mcp" in the first prompt - otherwise the agent will use Bash.

Simon Willison @simonw · July 2025 · via til.simonwillison.net

Part of "agentic manual testing": the agent must check its own code rather than assume it works.

Simon Willison @simonw · March 2026 · via simonwillison.net

"Almost all MCPs really should be CLIs" - dropped Playwright MCP in favor of chrome-devtools-mcp: context tax, the agent "ran Playwright where it could simply read the code".

Peter Steinberger @steipete · October 2025 · via steipete.me

Instead of 25+ tools, created playwrightess: one playwright_eval tool that runs JS in a persistent browser session.

Armin Ronacher @mitsuhiko · October 2025 · via lucumr.pocoo.org

"Constantly blows up the context window on nearly every call"; Microsoft confirmed this and recommends a CLI workflow for writing tests (issues #889, #1040).

Ask HN news.ycombinator.com · October 2025 · via news.ycombinator.com/item?id=45764043

6.0

Common issues and fixes

01“Executable doesn't exist” error / browser not found

Run npx playwright install chrome (or chromium). The Docker version downloads the browser itself.

02Client cannot see the server (0 tools)

Check in order: is the config JSON valid; is it the right config file (global vs project); fully restart the client. 90% of cases are one of these three.

03Windows: server does not start through npx

Wrap the command: "command": "cmd", "args": ["/c", "npx", "@playwright/mcp@latest"]. A classic Windows npx pitfall.

04Browser “busy”: profile is already in use

Close the running Chrome using the same profile or add --isolated (in-memory profile). Or connect to a live browser through --extension.

05Agent “goes blind”: page exists, but snapshot is empty/sparse

Wait for loading (browser_wait_for), check whether the content is in an iframe; for canvas/widget pages, enable --caps=vision and work from the screenshot.

06Consumes context: each step - tens of thousands of tokens

07Node < 18 or npx pulls an old version

Node 18+ is required. With NVM, specify the absolute path to node in command; @latest in args keeps the server current.

7.0

Frequently asked questions

▸What is Playwright MCP in simple terms?

▸How is it better than screenshot-based browser control?

▸How many tokens does it consume, and how can I reduce it?

▸How does Playwright MCP differ from Playwright itself?

▸Playwright MCP or Chrome DevTools MCP?

Playwright - browser driving (scenarios, forms, e2e); Chrome DevTools MCP - diagnostics (perf traces, network, console). Steve Kinney's formula: “driving vs debugging”. Both are often installed.

▸Is it safe?

▸Can I connect it to my already running Chrome?

Yes: the --extension flag connects the server to a live Chrome/Edge through the Playwright Extension, while the --cdp-endpoint flag connects to any browser via CDP.

▸Does it work from Russia?

Yes, completely: it is a local npm package, with no keys or payment required, and it does not send data externally. The only limitation comes from the client itself (Claude/Cursor), not the server.

▸Headless or with a window?

By default - with a window (you can see what the agent does). In CI/Docker, enable --headless; the Docker image supports headless chromium only.

alternatives

Chrome DevTools MCP →

official Chrome team server: stronger for debugging (profiling, DevTools protocol), weaker as “hands” for scenarios

ByteDance Browser MCP →

alternative agent browser implementation from the ByteDance monorepo

Browserbase / stealth forks

cloud browser and anti-bot bypasses - when scale or concealment is needed; this is no longer a “developer browser”

reg · io.github.microsoft/playwright-mcp

distribution · npm

card rev · 2026-07-02

data snapshot · July 2, 2026

Data: official MCP registry + GitHub API, daily snapshot. Score v0 - methodology · registry entry

Sections

14,522 servers

mcp catalog