MCP Accessibility Scanner 🔍

Star History

A powerful Model Context Protocol (MCP) server that provides automated web accessibility scanning and browser automation using Playwright and Axe-core. This server enables LLMs to perform WCAG compliance checks, interact with web pages, manage persistent browser sessions, and generate detailed accessibility reports with visual annotations.

Features

Accessibility Scanning

✅ Full WCAG 2.0/2.1/2.2 compliance checking (A, AA, AAA levels)
📄 Detailed JSON reports with remediation guidance
🎯 Support for specific violation categories (color contrast, ARIA, forms, keyboard navigation, etc.)

Browser Automation

🖱️ Click, hover, and drag elements using accessibility snapshots
⌨️ Type text and handle keyboard inputs
🔍 Capture page snapshots to discover all interactive elements
📸 Take screenshots and save PDFs
🎯 Support for both element-based and coordinate-based interactions

Advanced Features

📑 Tab management for multi-page workflows
🌐 Monitor console messages and network requests
⏱️ Wait for dynamic content to load
📁 Handle file uploads and browser dialogs
🔄 Navigate through browser history

Installation

You can install the package using any of these methods:

Using npm:

npm install -g mcp-accessibility-scanner

Installation with Docker

A pre-built image is available on Docker Hub. The image includes Chromium and is pre-configured for containerized use — no extra flags needed.

Pull from Docker Hub:

docker pull justasmonkev/mcp-accessibility-scanner

Claude Code

 claude mcp add mcp-accessibility-scanner -s user -- docker run -i --rm justasmonkev/mcp-accessibility-scanner

To persist screenshots and reports on your host, add a volume mount:

claude mcp add mcp-accessibility-scanner -s user \
  -- docker run -i --rm -v /tmp/mcp-output:/app/output justasmonkev/mcp-accessibility-scanner

Without the -v mount, output files only exist inside the container and are lost when it exits.

Docker Compose

docker compose up -d

The Compose configuration publishes the unauthenticated MCP HTTP transport on 127.0.0.1:8931 only. Do not expose this port to untrusted networks.

Build from source

docker build -t mcp-accessibility-scanner .

Docker smoke test

npm run test:docker

Installation in VS Code

Install the Accessibility Scanner in VS Code using the VS Code CLI:

For VS Code:

code --add-mcp '{"name":"accessibility-scanner","command":"npx","args":["mcp-accessibility-scanner"]}'

For VS Code Insiders:

code-insiders --add-mcp '{"name":"accessibility-scanner","command":"npx","args":["mcp-accessibility-scanner"]}'

CLI Modes

The scanner can run in two modes depending on how you use it.

MCP server (default, no subcommand)

When launched without a subcommand, the process starts an MCP server that communicates over stdio. This is the mode used by MCP clients such as Claude Desktop, VS Code, and Claude Code -- you should never need to run it by hand.

npx mcp-accessibility-scanner            # starts the MCP server (stdio)

All of the MCP client configuration examples in this README already use this default mode.

Interactive REPL (interactive subcommand)

For manual terminal use, the interactive subcommand starts a readline REPL where you can call any tool directly:

$ npx mcp-accessibility-scanner interactive
Interactive mode. Type "<tool-name> <json>" to call a tool. Ctrl+D to exit.
> browser_navigate {"url": "https://example.com"}
> scan_page {"violationsTag": ["wcag21aa"]}
> audit_keyboard {"maxTabs": 30}
> audit_screen_reader {}

Each line is <tool-name> <json-arguments>. Omit the JSON to pass {}. Global browser connection flags still apply here, for example npx mcp-accessibility-scanner --headless interactive. Use --mobile or PLAYWRIGHT_MCP_MOBILE=1 to emulate a generic mobile device (Pixel 10 for Chromium, iPhone 17 for WebKit). It cannot be combined with --device, CDP attach/launch modes, remote browser endpoints, or --extension.

Browser extension mode

Use --extension to connect through the current Playwright Extension, which must support extension protocol v2.

npx mcp-accessibility-scanner --extension

Set PLAYWRIGHT_MCP_EXTENSION_TOKEN to the token shown by the extension to bypass the connection approval dialog. The relay's CDP WebSocket endpoint always requires a separate random token, generated per relay and appended automatically for the server's own connection. This CDP token is never passed in Chrome's launch arguments or extension URL; the extension approval token cannot authenticate a CDP client. Token-bypass connections are not background-safe: Chrome focuses the connection tab and window, and client-created tabs remain open after disconnect (upstream limitation). With a token, the extension must connect and finish setup within 30 seconds after the connection page opens. Failed attempts release the relay so the next tool call can retry. Without a token, manual approval waits until you approve or cancel the call. When --user-data-dir contains multiple Chrome profiles, the profile with the extension installed is selected automatically, preferring Chrome's last-used profile. Pass --profile-dir-name (or set PLAYWRIGHT_MCP_PROFILE_DIR_NAME) to select a profile explicitly; it requires --user-data-dir and accepts a Chrome profile directory name such as Default or Profile 1 (see "Profile Path" at chrome://version). The PLAYWRIGHT_MCP_EXTENSION_TOKEN approval token is specific to the profile, so when selecting a profile explicitly, use the token shown in that profile.

Packed extensions require an enabled record in the profile's preferences; a leftover extension directory alone does not count as installed. Profiles whose preferences mark the extension disabled or uninstalled are excluded from automatic selection and rejected for explicit selection. Explicit selection with a custom executable skips local installation checks. A whitespace-only PLAYWRIGHT_MCP_PROFILE_DIR_NAME is treated as unset; other string environment variables retain their existing blank-value handling.

Discovering available tools (list-tools subcommand)

To print every tool name and its description:

npx mcp-accessibility-scanner list-tools

Note: Tool names like browser_navigate and scan_page are MCP tool identifiers (and REPL commands in interactive mode). They are not shell subcommands -- you cannot run npx mcp-accessibility-scanner browser_navigate.

Configuration

Here's the Claude Desktop configuration:

{
  "mcpServers": {
    "accessibility-scanner": {
      "command": "npx",
      "args": ["-y", "mcp-accessibility-scanner"]
    }
  }
}

Advanced Configuration

You can pass a configuration file to customize Playwright behavior:

{
  "mcpServers": {
    "accessibility-scanner": {
      "command": "npx",
      "args": ["-y", "mcp-accessibility-scanner", "--config", "/path/to/config.json"]
    }
  }
}

Configuration Options

Create a config.json file with the following options:

{
  "browser": {
    "browserName": "chromium",
    "launchOptions": {
      "headless": true,
      "channel": "chrome"
    },
    "cdpLaunch": {
      "command": "open",
      "args": ["-a", "Slack", "--args", "--remote-debugging-port={port}"],
      "startupTimeoutMs": 30000
    }
  },
  "timeouts": {
    "navigationTimeout": 60000,
    "defaultTimeout": 5000,
    "settle": 500,
    "idle": 0
  },
  "network": {
    "allowedOrigins": ["example.com", "trusted-site.com"],
    "blockedOrigins": ["ads.example.com"]
  },
  "snapshot": {
    "boxes": true
  }
}

Available Options:

• browser.browserName: Browser to use (chromium, firefox, webkit)

• browser.allowedUploadDirs: Restrict files sent by browser_file_upload and browser_drop to regular files inside these directories, including resolved symlink targets. Restricted uploads and drops use a checked file handle and accept up to 50 MiB total per call. Unset allows any path; [] denies all file uploads and drops (text-only drops still work). Blank list entries are rejected. CLI: --allowed-upload-dirs (semicolon-separated; "" denies all), env: PLAYWRIGHT_MCP_ALLOWED_UPLOAD_DIRS (empty string denies all).

  The list must be an array, not null. Roots must exist at startup: their canonical paths are resolved once and retained for the server's lifetime, so retargeting a configured symlink does not grant access to a new tree. Non-empty upload allowlists require macOS or Linux with /proc/self/fd available. macOS blocks ancestor symlinks during the file open; Linux checks the opened descriptor's path. Other platforms reject restricted file uploads and drops rather than rely on race-prone pathname checks. Unrestricted uploads, deny-all lists, and text-only drops keep working on all platforms.

• browser.launchOptions.headless: Run browser in headless mode (default: true on Linux without display, false otherwise)

• browser.launchOptions.channel: Browser channel (chrome, chrome-beta, msedge, etc.)

• browser.contextOptions.clientCertificates: With client certificates, a configured launch proxy is also applied to fresh contexts unless browser.contextOptions.proxy explicitly overrides it. Playwright 1.63.0 ignores certificate-interceptor proxy bypass rules, so combining certificates with a nonblank effective proxy.bypass is rejected before launch. For a fresh context on a remote browser, configure its proxy explicitly; a proxy inherited only from that browser cannot be inspected. Existing attached contexts retain their current settings.

• browser.launchOptions.chromiumSandbox: Defaults to false for downloaded Chromium builds on Linux because they lack the setuid sandbox helper, and true otherwise. Remote and VS Code endpoints choose on the remote host. An explicit config or PLAYWRIGHT_MCP_SANDBOX value wins; --no-sandbox always disables it.

• browser.cdpEndpoint: Attach to an already-running Chromium-family app with CDP enabled

• browser.cdpHeaders: Map of HTTP headers to send with the CDP connect request, e.g. { "Authorization": "Bearer <token>" }, for endpoints that require header-based authentication

• browser.cdpTimeout: Maximum time in milliseconds to wait when connecting to the CDP endpoint (default: 30000)

• browser.cdpLaunch: Launch a Chromium-family desktop app with CDP enabled, wait for the endpoint, and manage the child process lifecycle

• CDP attach modes preserve the target browser's existing default-context settings instead of applying Playwright's defaults.

• browser.contextOptions.storageState: Start a fresh context from a recorded Playwright storage state. Imports into existing CDP or VS Code contexts are rejected; use --isolated for CDP or sign in interactively. The default persistent mode uses a fresh disposable profile. See Auditing pages behind a login.

• browser.profileDirName: Chrome profile directory name used in extension mode, for example Default or Profile 1 (CLI: --profile-dir-name, env: PLAYWRIGHT_MCP_PROFILE_DIR_NAME). Requires --user-data-dir and extension mode (--extension or --connect-tool); defaults to the last-used profile that has the extension installed.

• timeouts.navigationTimeout: Maximum time for page navigation in milliseconds (default: 60000)

• timeouts.defaultTimeout: Default timeout for Playwright operations in milliseconds (default: 5000)

• timeouts.settle: How long to wait after every action before responding (default: 500). An action that finishes quietly is first watched for up to 100ms (or the settle delay, whichever is shorter) so scheduled network work can still be awaited before the settle delay.

• timeouts.idle: Release the default browser context after this many idle milliseconds (default: 0, disabled). Accepts integers from 0 to 2147483647.

• network.allowedOrigins: List of origins to allow (blocks all others if specified)

• network.blockedOrigins: List of origins to block

• snapshot.boxes: Include each element's viewport-relative bounding box as [box=x,y,width,height] in snapshots (default: false; CLI: --snapshot-boxes, env: PLAYWRIGHT_MCP_SNAPSHOT_BOXES=1)

• imageResponses: allow (default) returns text and images; omit excludes images; only omits text from successful responses containing images. Errors, browser lifecycle notices, and responses without images (including full-page screenshots) keep their text. Structured results and resource links are always preserved. In only mode, screenshot save-path text, generated code and any accompanying text-only findings are omitted; use allow if you need them. Interactive mode rejects only because its REPL prints text only; use allow or omit there. auto remains a legacy alias for allow. CLI: --image-responses only; env: PLAYWRIGHT_MCP_IMAGE_RESPONSES=only. Precedence: CLI, then environment, then config file.

• server.authToken: When set, Streamable HTTP requests (--port) require Authorization: Bearer <token> or return 401 (env: PLAYWRIGHT_MCP_AUTH_TOKEN). Blank or malformed tokens fail at startup. The scheme is case-insensitive; the token is exact. Bearer auth does not encrypt traffic: authenticated listeners must bind to loopback, such as --host 127.0.0.1; use a TLS reverse proxy for remote access. The printed client config includes a header placeholder to replace locally, without logging the secret. Unset keeps unauthenticated access.

• filePaths: Optional relative or absolute rendering for output paths in screenshot/PDF results and generated code, report metadata, annotated/keyboard screenshots, and download summaries. Relative paths are based on the server working directory, not the client workspace or output directory. File writes, stored report contents, and absolute file:// resource URIs are unchanged. Omitted preserves legacy rendering: configured relative output directories produce relative paths, while default temp output paths are absolute (unlike upstream Playwright’s relative default). CLI: --file-paths absolute; env: PLAYWRIGHT_MCP_FILE_PATHS=absolute. Precedence: CLI, then environment, then config file.

• outputDir: Directory for output files — reports, screenshots, traces, and session logs (CLI: --output-dir, env: PLAYWRIGHT_MCP_OUTPUT_DIR). Defaults to a fresh directory under the system temp folder, resolved once per server run so all of a run's artifacts land together. The output location is always server configuration; the deprecated MCP roots capability (client workspace folders) is no longer consulted.

CLI equivalents are also available: --cdp-launch-command, --cdp-launch-args, --cdp-launch-cwd, --cdp-launch-port, --cdp-launch-startup-timeout, --cdp-endpoint, --cdp-header (repeat for multiple headers, e.g. --cdp-header "Authorization: Bearer <token>"), and --cdp-timeout. The CDP headers and timeout can also be set via the PLAYWRIGHT_MCP_CDP_HEADERS (one Name: Value entry per line) and PLAYWRIGHT_MCP_CDP_TIMEOUT environment variables.

If CDP attachment times out after the WebSocket connects, an existing sleeping or unresponsive tab may be blocking Playwright's browser initialization (upstream report). Use an explicit positive --cdp-timeout to bound the attempt. Inspect or wake the affected tabs yourself, or attach to a separate disposable browser. noDefaults and --isolated do not skip initialization of existing tabs; the server does not close your tabs or bypass Playwright's initialization to work around this.

Playwright 1.63 does not support back/forward-cache (BFCache) restoration: an attached browser with BFCache enabled can return unusable references and omit iframe contents after back navigation. This was reproduced with full Chromium, Chrome, and Edge. For a browser you launch yourself, include --disable-back-forward-cache before attaching, or use the server's normal browser launch mode, which retains Playwright's default flag. The server does not change an attached browser's flags or replace back navigation with a reload. See the upstream maintainer's explanation, Playwright's BFCache limitation, and the verification and reproduction command.

For remote HTTP access, configure the TLS reverse proxy explicitly. For example, with the MCP server bound using --host 127.0.0.1 --port 8931 and PLAYWRIGHT_MCP_AUTH_TOKEN set:

• Accept only your configured public hostname over HTTPS and forward /mcp to http://127.0.0.1:8931/mcp.
• Set the upstream Host header to 127.0.0.1:8931, not the public hostname. Forward the client's Authorization header unchanged; do not inject a shared token for unauthenticated clients.
• Before removing Origin, reject any non-empty value outside your explicit trusted HTTPS origin list (for example, https://mcp.example.com). Allow absent Origin for non-browser clients. Then remove Origin upstream, or rewrite it to http://127.0.0.1:8931. Never strip arbitrary origins without checking them first.
• Disable response buffering for SSE streams. Browser clients on a different origin also need a narrowly scoped CORS policy at the proxy.

The server does not trust Forwarded or X-Forwarded-* to bypass its checks. Preserving the public Host or HTTPS Origin upstream returns 403, even with a valid bearer token.

Caller-supplied screenshot, PDF, scan-page-matrix, and audit report filenames use a no-clobber policy: an existing file causes the tool call to fail instead of being overwritten. Windows-reserved basenames and names ending in a dot or space are rejected on every platform so configured names behave consistently across hosts.

Failed download saves are reported as tool errors in the current or next response, including after the tab closes; failed entries do not advertise a saved file or an ongoing download. The server retains up to 20 bounded error messages between responses and reports any omitted count. Full Chromium-family browsers on Playwright 1.63.0 have verified native crashes after persistent-profile relaunch. Error reporting does not fix that native crash. The explicit --isolated control passed, but does not preserve a profile between launches; use recorded storage state when needed for authentication. Existing profiles and browser defaults are preserved.

Use --timeout-settle or PLAYWRIGHT_MCP_TIMEOUT_SETTLE to override the post-action settle delay. It applies after every action so delayed DOM-only updates are included in the response; a short observation window also catches scheduled requests and waits for them before that delay.

Use --timeout-idle 300000, timeouts.idle, or PLAYWRIGHT_MCP_TIMEOUT_IDLE to release the default browser after five idle minutes. Shared contexts stay open while any client is working; the idle window starts after the last tool call or download finishes. Explicit recordings prevent idle release until browser_stop_recording finishes; passive --save-session capture does not. Cleanup finalizes traces. The next browser tool call reopens the connection and includes a note to navigate again and refresh element references. Attached CDP, extension, and VS Code browsers are disconnected; their external pages remain open. Close and session-management tools do not relaunch an idle browser. Explicit browser_session_open handles keep their separate PLAYWRIGHT_MCP_BROWSER_SESSION_TTL_MS behavior. Zero disables this feature; blank environment values leave the existing configuration unchanged.

The VS Code browser_connect tool accepts only playwright or playwright-core libraries and loopback WebSocket URLs. Set PLAYWRIGHT_MCP_VSCODE_ALLOW_REMOTE=1 to allow remote endpoints, which must use wss:. URL userinfo credentials are rejected.

HTTP Heartbeat

When the server runs with --port, it sends MCP heartbeat pings after a Streamable HTTP client opens the optional event stream. POST-only clients stay connected without heartbeat because server-initiated requests cannot reach them. Set PLAYWRIGHT_MCP_PING_TIMEOUT_MS to override the default 5000 ms timeout, or to 0 or any negative value to disable heartbeat pings. A client that answers ping with a JSON-RPC "method not found" error is treated as alive: the server stops heartbeating that session instead of closing it. Only an unanswered ping (timeout) or a transport failure closes the session.

Clients without the initialize handshake

Clients on the MCP 2026-07-28 revision no longer send the initialize handshake. With --port, requests carrying the revision's per-request _meta envelope are served natively on the 2026-07-28 protocol: server/discover is answered (so clients negotiating with versionNegotiation: 'auto' or a 2026-07-28 pin connect directly), results carry resultType and the SEP-2549 cache fields — the tool list is advertised as cacheable for one hour with cacheScope: "private" — and the SEP-2243 standard headers (MCP-Protocol-Version, Mcp-Method, Mcp-Name) are validated against the request body. Older handshake-free clients (2025-era requests without the envelope) are served statelessly as before. In both cases requests receive no heartbeat pings, and in the modes where the server creates browser contexts itself each request runs against a fresh default browser session: with the default persistent profile the per-request default context runs in its own disposable profile (like an explicit browser session), so parallel handshake-free requests do not contend for the stable profile — and the stable profile's sign-in state is not visible to them — while --isolated, remote endpoints and isolated CDP modes mint a fresh context per request anyway. Modes that reuse one live browser context are the exception: --extension (and a browser_connect or VS Code session switched to a connected-browser provider) and CDP attach without --isolated serve every handshake-free request from the same shared context, so its tabs, cookies and storage persist across requests — the same sharing that makes these modes refuse browser_session_open (in --vscode serving the session tools are the exception: they are host-scoped and keep running against the default provider even while switched — see Browser Session Tools). With a pinned --cdp-launch-port, only one launched application can be served at a time, so a second handshake-free request arriving while another request's browser context is still live is rejected with a clear error instead of silently attaching to the first request's application. With --user-data-dir, each handshake-free request launches a browser in the one configured profile: the profile's state persists across requests, and parallel requests contend for its browser lock and can fail with "Browser is already in use". Elsewhere, browser state that must persist across handshake-free requests belongs in an explicit browser session — a browserSessionId handle minted by browser_session_open in one request resolves in later ones (see Browser Session Tools). Clients that do send initialize keep the classic Mcp-Session-Id session behavior unchanged. When several such stateful clients are connected at once in the default persistent-profile mode, the first client's default context holds the stable profile — concurrent clients' default contexts run in their own disposable profiles (without the stable profile's sign-in state) until it is freed, instead of failing with "Browser is already in use".

Auditing pages behind a login

Most real audits target pages that only exist for a signed-in user. There are two ways to get there.

Interactive route (no setup)

Every tool shares one browser context, and audit_site crawls in a temporary tab of that same context, so cookies and local storage created while you drive the browser are already available to the crawl:

1. browser_navigate to the login page
2. browser_fill_form / browser_click to sign in
3. browser_navigate to the first page you want audited
4. audit_site — the crawl inherits the session you just created

This works out of the box in every mode, including the default persistent-profile mode. With the default profile the session also survives across server restarts, so you usually only sign in once. The default profile is keyed to the server's working directory, so each workspace's server keeps its own sign-in state — servers launched for different workspaces neither share cookies nor contend for the same profile.

Storage state route (repeatable, CI-friendly)

Record a session once with Playwright's codegen, then hand the file to the server:

npx playwright@1.63.0 codegen --save-storage=auth.json https://example.com/login

Sign in in the opened browser, then close it — auth.json now holds the cookies and local storage.

Pass it to the server with the CLI flag, the environment variable, or the config file:

npx mcp-accessibility-scanner --isolated --storage-state ./auth.json

PLAYWRIGHT_MCP_ISOLATED=true PLAYWRIGHT_MCP_STORAGE_STATE=./auth.json npx mcp-accessibility-scanner

{
  "browser": {
    "isolated": true,
    "contextOptions": {
      "storageState": "./auth.json"
    }
  }
}

Every supported mode handles the state — by applying it or refusing it.

Playwright 1.63.0 safety restriction: importing into an existing context is rejected before taking a rollback snapshot or resetting any storage. On this pin, snapshot capture can execute service-worker-served scripts for a previously visited origin whose tab is no longer open (upstream fix). Use a fresh context, or omit --storage-state and sign in interactively. Service workers are not disabled. A future dependency upgrade must also pass the recorder/shared-client checks in #218 and IndexedDB checks in #224 before this restriction is reconsidered.

IndexedDB snapshot limitation: on pinned Playwright 1.63.0 with Chromium 153.0.8010.12 and Firefox 155.0, storageState({ indexedDB: true }) loses Map and Set contents. Both newContext({ storageState }) and setStorageState() restore them as empty plain objects; ordinary JSON records survive. Fresh contexts protect existing browser data, but cannot recover values already lost during capture. The upstream fix is merged but is not in this pin. Before allowing imports into existing contexts again, verify both restore paths preserve Map/Set types and entries on each supported engine, including after a failed import. The real-browser regression in tests/browser-failures.integration.test.ts checks that rejecting an import leaves the original Map/Set records intact and that isolated JSON IndexedDB imports still work.

• Fresh-context modes (--isolated, the remote-endpoint mode, or either CDP mode combined with --isolated): the context is created with the storage state directly.
• Default persistent-profile mode with --storage-state: the session runs in a fresh, disposable profile — unique to that session and removed when it closes — built from the state, so the recorded state is provably the only session data (without --storage-state the regular persistent profile is used and survives restarts, as before). Any page the launch opened (for example from a URL in browser.launchOptions.args) is parked on a blank replacement before the state lands, then the replacement is navigated to the same URL, so a still-running anonymous page cannot overwrite the recorded identity and a scan never reads its DOM. This also means --storage-state cannot be combined with --user-data-dir (a user-supplied profile carries its own session and will not be wiped; the server refuses the combination).
• CDP modes without --isolated and the VS Code provider: --storage-state is rejected when the browser already has a context. Add --isolated in CDP mode to create a fresh context; otherwise omit the state and sign in interactively. If the browser exposes no context, the server creates one with the state. CDP sessions joining that same server-created context inherit its live state without resetting it.
• --extension (with or without --isolated) refuses storage imports entirely: it works through the browser you are already running, where wiping cookies to install a recorded state is not an acceptable side effect, so the server refuses to start rather than doing that silently. There, sign in interactively instead — the persistent profile also keeps the session across restarts.

Keep the crawl from destroying its own session

audit_site excludes logout|signout by default, which is not enough for most applications. Add anything else that ends or changes the session before you start the crawl:

{
  "excludePathPatterns": ["logout|signout", "account/(close|delete)", "sessions/revoke", "/switch-(locale|account|org)"]
}

Note that excludePathPatterns replaces the default rather than extending it, so repeat logout|signout in your list.

If a session cookie disappears anyway, audit_site says so instead of reporting a confident, wrong audit: the result starts with a WARNING: cookie(s) … disappeared while loading <url> line, and both the JSON report and the structured content carry a sessionLosses list naming, for each lost cookie, the page that dropped it — the page reached after any redirect, and reported even when that page failed to finish loading. If one of the lost cookies was the session, every page scanned after that point was audited as a signed-out user — exclude the offending URL, sign in again, and re-run.

The check compares which cookies the crawled URLs carry, not their values, so a rotating CSRF token never reads as a lost session. A cookie the browser deleted at its own stated expiry is ignored for the same reason — Cloudflare's __cf_bm lives 30 minutes and would otherwise warn on any longer crawl. Beyond that no attempt is made to tell an authentication cookie from any other: nothing in a cookie marks it as one, so any cookie the crawl started with and later lost is reported. Monitoring does not stop at the first loss — each cookie is reported once, at the URL where it vanished, so an analytics cookie expiring early cannot mask the session cookie being dropped later. URLs discovered mid-crawl join the cookie tracking before they are visited, so a session cookie scoped to a path below the start URL (say /app) is watched too.

Available Tools

Page-registered WebMCP tools are not currently exposed. See the WebMCP adoption decision for the deferral and conditions for revisiting an opt-in capability.

The MCP server provides comprehensive browser automation and accessibility scanning tools:

Core Accessibility Tool

scan_page

Performs a comprehensive accessibility scan on the current page using Axe-core.

Parameters:

• violationsTag: Array of WCAG/violation tags to check
• includeIncomplete (default true): also report Axe "incomplete" results
• maxNodesPerViolation (default 10): cap on nodes reported per rule
• includeSelectors / excludeSelectors: CSS selectors that scope the scan
• withRules / disableRules: Axe rule ids that narrow which rules run
• annotateScreenshot (default false): capture an annotated screenshot of the violations

Annotated screenshots: When annotateScreenshot is true, each violating element is outlined and labelled with the rule ids it failed, a full-page PNG is written to the MCP output directory (scan-page-annotated-{timestamp}-{token}.png) and returned as a resource link, and the markers are then removed so the page is left exactly as it was. The markers are drawn in an out-of-flow overlay clipped to each element's own box, so they never reflow the page. The overlay uses a fresh id per scan, is placed in the browser's top layer so it stays visible over an open dialog, popover or fullscreen element, and compensates for a CSS zoom or a scaled ancestor so markers line up with what is rendered. An element that fails several rules gets one box listing every rule id, and elements inside open shadow roots are marked by walking the shadow path Axe reports. Running animations are frozen at their current time before the elements are measured and resumed after the capture, so a moving target keeps its marker. The markers themselves live in a shadow root under an overlay whose own styles are !important, so page CSS cannot restyle or hide what the report counts, and each rule label sits outside the clipped box so it stays readable on an element smaller than its own label. At most 50 elements are annotated per scan. The result text always reports how many nodes were marked out of the total, plus how many were left out because they exceeded the limit, were hidden, zero-size or off-canvas (a full-page screenshot is clipped to the document box), or were inside an iframe (cross-frame selectors cannot be resolved from the top document).

Supported Violation Tags:

• WCAG standards (in the default set): wcag2a, wcag2aa, wcag2aaa, wcag21a, wcag21aa, wcag21aaa, wcag22a, wcag22aa, wcag22aaa
• Section 508 (in the default set): section508
• Categories (opt-in): cat.aria, cat.color, cat.forms, cat.keyboard, cat.language, cat.name-role-value, cat.parsing, cat.semantics, cat.sensory-and-visual-cues, cat.structure, cat.tables, cat.text-alternatives, cat.time-and-media
• Non-conformance tags (opt-in): best-practice, experimental (see the caveat below -- a few experimental rules also carry a WCAG tag and run by default)

The default set is the WCAG and Section 508 tags only, so a default report means "this fails a conformance criterion". Category tags are opt-in for that reason: Axe matches requested tags with OR, so asking for cat.keyboard also pulls in best-practice rules such as region and skip-link that carry both tags. No live conformance rule is lost by leaving them out: the only rules reachable only through a cat.* tag are duplicate-id and duplicate-id-active, which Axe marks deprecated because WCAG removed SC 4.1.1. Add best-practice (landmark structure, heading order, tabindex hygiene) or a cat.* tag when you want that broader review.

The same OR semantics apply to experimental, with one deliberate exception: five experimental rules -- css-orientation-lock (SC 1.3.4), label-content-name-mismatch (SC 2.5.3), p-as-heading, table-fake-caption and td-has-header (SC 1.3.1) -- also carry a wcag* tag and so run in the default set. In Axe, experimental describes how settled the heuristic is, not whether the criterion is real, so these are kept rather than filtered out. Adding the experimental tag pulls in the remaining experimental rules, which have no conformance tag of their own.

Scan scoping: scan_page, audit_site, and scan_page_matrix accept includeSelectors and excludeSelectors to limit what Axe looks at. Use includeSelectors to audit one component (["#checkout-form"]) and excludeSelectors to drop third-party noise that pollutes every report (["#cookie-banner", "iframe.intercom-frame"]). Exclusions are applied after inclusions, so you can carve a widget out of an included subtree.

Selectors are resolved before the scan runs:

• Syntactically invalid CSS fails the scan, naming the selector.
• An includeSelectors entry that matches nothing fails the scan. Axe on its own would accept a partly-matching include set and quietly scan less than you asked for, so the scanner refuses rather than returning a clean-looking report with half the scope missing.
• An excludeSelectors entry that matches nothing is a no-op, not an error -- a crawl legitimately visits pages that lack the excluded widget.

In audit_site, selectors apply to every crawled page, so an includeSelectors value that is absent from a given page marks that page as errored in the report while the crawl continues. Link discovery runs before the scan, so pages reachable only through an errored page are still crawled.

Rule-level control: scan_page, audit_site, and scan_page_matrix accept withRules and disableRules to pick individual Axe rules instead of whole tag sets. Use withRules to re-check one rule after a fix (["color-contrast"]) and disableRules to mute a rule you have already triaged (["region"]). Rule ids are the ones Axe reports (image-alt, color-contrast, ...); see the Deque rule reference.

• withRules overrides violationsTag. Axe can run either a rule list or a tag list, never both, so when withRules is set the tags are ignored entirely -- withRules: ["image-alt"] runs exactly that one rule regardless of violationsTag. Rule ids are the more specific request, so they win.
• disableRules subtracts from whatever is selected. It applies to violationsTag and withRules alike. (Axe itself ignores disabled rules once you give it an explicit rule list; the scanner subtracts them up front so the two options mean the same thing together as apart.) Disabling every rule in withRules is an error rather than an empty scan.
• An explicitly empty withRules is an error too. withRules: [] selects no rules, and silently falling back to the tag set would run a different scan than the one requested — omit the option to scan by tags instead. Clients that build the list dynamically should drop the key when the list comes out empty.
• Unknown rule ids fail the scan, naming the id. Both options are checked against Axe's rule catalogue before the browser is touched, so a typo is reported as Unknown Axe rule id(s) in withRules: image-altt rather than surfacing later as an frame.evaluate failure from inside the page. Rule ids apply to a whole run, so audit_site and scan_page_matrix check them once before they touch the page -- a bad id fails the call outright instead of crawling every URL, or reloading and re-emulating the page, before rejecting the argument.

audit_site and scan_page_matrix record both values in their JSON report metadata, so a stored report can be told apart from a full scan.

Incomplete ("needs review") results: Axe returns incomplete for checks it cannot decide on its own -- contrast over a background image or gradient, ambiguous labels, elements it could not fully evaluate. scan_page, audit_site, and scan_page_matrix report these in a section separate from violations so you can resolve them by inspecting the page (screenshot, snapshot, browser_evaluate). Set includeIncomplete: false to suppress them.

Frames that could not be scanned: Axe is installed into every frame of the page before the scan runs. A frame that navigates mid-injection, or whose renderer does not answer within a second, is left out -- and its contents then contribute no findings. Rather than let that pass as a clean result, all three scan tools print a WARNING: Axe could not be installed in N frame(s) block listing the frame URLs, and audit_site and scan_page_matrix also record them per page and per variant in their JSON reports (unscannedFrames) and in structuredContent. A frame that was still loading usually succeeds on a re-run; one that fails consistently has to be audited on its own.

A nested frame is reported when any frame above it went unscanned, even if its own injection succeeded: Axe reaches a nested document only by relaying through the frames above it, so an outer frame without Axe takes everything below it out of the scan.

A frame you scoped out yourself is not reported: with excludeSelectors: ["iframe.intercom-frame"] that widget failing to load is the outcome you asked for, not a gap. Scope is resolved through the whole frame chain and across shadow boundaries, so an includeSelectors entry naming an ancestor still covers frames nested several levels below it or inside a shadow root, and excluding an outer frame or a shadow host silences everything inside it. Anything the check cannot resolve is reported rather than hidden.

Audit Tools

audit_site

Crawls and scans multiple internal pages, then aggregates violations across the site.

• Default strategy: link-based BFS from the current URL
• Supports links, nav, sitemap, and provided URL strategies
• Sitemap URLs and every redirect must pass the server network policy and crawl scope. Fetches run on the MCP host, use HTTP(S) without browser cookies or auth headers, and have a 15-second total timeout, 20-redirect cap, and 10 MiB response limit. Browser proxy settings, browser.remoteEndpoint, browser.cdpEndpoint (including loopback endpoints, which may tunnel to remote browsers), and switched browser_connect providers are rejected for this strategy; use provided URLs in these modes. Sitemap TLS certificates must be valid even when browser HTTPS errors are ignored.
• Always writes a JSON report (default filename: audit-site-{timestamp}-{token}.json)
• Warns and records sessionLosses if the crawl loses cookies it started with — see Auditing pages behind a login

Example flow:

1. Navigate to your site homepage with browser_navigate
2. Run audit_site with maxPages: 25 and maxDepth: 2
3. Review the report path returned by the tool (written to the MCP output directory)

scan_page_matrix

Runs Axe scans on the same page across viewport/media/zoom variants and compares deltas against baseline.

• Default variants: baseline, mobile, desktop, forced-colors, reduced-motion, zoom-200
• Supports custom variants and optional reload between variants
• Always writes a JSON report (default filename: scan-matrix-{timestamp}-{token}.json)
• JSON report and structured result schema v2 set baseline deltas to null when either scan left frames unscanned, because their coverage is not comparable

Example flow:

1. Navigate to a page state you want to validate
2. Run scan_page_matrix with defaults (or provide custom variants)
3. Review per-variant deltas and open the generated JSON report path

audit_keyboard

Audits real keyboard focus behavior by pressing Tab (and optional Shift+Tab) with practical heuristics.

• Checks skip links, focus visibility, focus jumps, and possible focus traps
• Checks target size against WCAG 2.2 SC 2.5.8 (checkTargetSize, default on)
• Checks that focus is not entirely obscured, WCAG 2.2 SC 2.4.11 (checkFocusObscured, default on)
• Optional issue screenshots (screenshotOnIssue)
• Always writes a JSON report (default filename: audit-keyboard-{timestamp}-{token}.json)

Limits of the WCAG 2.2 checks — these are heuristics, not a conformance verdict:

• Target size only inspects elements the tab order actually reaches, so pointer-only targets are never measured.
• Of the SC 2.5.8 exceptions, only spacing (a 24px-diameter circle centered on the target must reach neither another target's box nor another undersized target's circle) and inline (an inline-level target — inline, inline-block, inline-flex, … — inside surrounding sentence text, found by walking out through inline wrappers such as <strong> to the containing block) are evaluated. The user agent control, essential, and equivalent exceptions cannot be detected from the DOM, so a target relying on one of them is still reported and needs manual triage.
• Spacing neighbours use the same pointer-target rule as the focused element, so rendered :disabled controls are not counted as neighbours, and contenteditable regions are counted as targets on both sides.
• Target size uses the element's bounding box, so an inline target wrapped over several lines is measured as one union box rather than per line, and a target whose visible area is cut down by an overflow or clip-path ancestor is measured at its full unclipped size.
• SC 2.4.11 is the Minimum (AA) level: a focused element is only reported when every sampled point of its box is covered by other content. Partially covered focus passes here, and the stricter SC 2.4.12 (AAA) is not checked. It applies to every focus stop with a rendered box, including elements that are not pointer targets such as iframes.
• Coverage is measured by hit-testing sample points and then checking that the element hit actually paints (visible, non-zero opacity all the way up to the first wrapper shared with the focused element, non-transparent background or background image). A transparent click-catching overlay therefore does not count as obscuration, but a covering layer with pointer-events: none is never returned by hit testing and is missed. Semi-transparent overlays that still leave content legible are reported.

Example flow:

1. Navigate to the target page and let it fully load
2. Run audit_keyboard with maxTabs: 50
3. Review focus findings and open the generated JSON report path

audit_screen_reader

Audits what a screen reader actually announces, using the browser's own accessibility tree (page.ariaSnapshot) plus element geometry. No screen reader is installed or driven; this is a static reading of the exposed tree.

Checks (checkNames)

• missing-accessible-name: controls and images exposed with no accessible name (WCAG 4.1.2)
• uninformative-accessible-name: names such as "click here", "read more", "image" that mean nothing out of context (WCAG 2.4.4)
• filename-as-accessible-name: image alt text that is a file name, e.g. IMG_1234.jpg, DSC00123 (WCAG 1.1.1). Only images are checked: a link or button legitimately named after the file it downloads (logo.png) is not a defect.
• label-in-name-mismatch: the accessible name does not contain the visible label, which breaks voice control (WCAG 2.5.3). The visible label of <input type="submit|button|reset"> is read from its value, and a web component's label is read from its open shadow root.
• duplicate-accessible-name: sibling links with the same name that lead to different URLs (WCAG 2.4.4)

Check (checkReadingOrder)

• reading-order-mismatch: accessibility tree order (what is read) versus visual position (WCAG 1.3.2), i.e. order, flex-direction: row-reverse, absolute positioning

What it deliberately does not detect

• Reading order is only compared between siblings that form a single row or a single column. Genuine two-dimensional layouts (grid, CSS multi-column, wrapped flex) have no single correct linear order and are skipped rather than guessed.
• Elements are excluded from the reading-order comparison when they render no text, are aria-hidden, floated, position: fixed, off-canvas, or clipped to 1px, because their visual position is decoupled from source order by design. Tolerance: two boxes count as swapped only when they are fully separated along the compared axis (1px), and right-to-left containers are compared right-to-left.
• Duplicate names are only reported when the destinations differ and are observable, which today means resolved link URLs (/help and https://site/help are the same destination). Two Save submit buttons in one form are never called ambiguous, because nothing in the exposed tree says whether they do different things.
• Only elements the AI snapshot gives a ref are analyzed, and Playwright refs the elements that are visible and receive pointer events. A control that is announced but not interactable (pointer-events: none, some off-canvas widgets) is therefore skipped: without a ref it cannot be measured, so neither its aria-hidden state nor a selector to fix it can be established, and reporting it would mostly surface decorative aria-hidden icons.
• Heading levels and landmark structure are not checked; axe already reports those (heading-order, region, landmark-one-main), so use scan_page for them.
• Findings for names overlap with axe rules such as link-name, button-name and image-alt; this tool adds the quality checks (generic names, file names, label-in-name, duplicates) that axe cannot make.
• It reports the page as currently rendered. Content behind a collapsed panel or another viewport is judged in that state.

How names are measured: the AI snapshot omits a name that a control's rendered children carry (<a><strong>Docs</strong></a>, a button whose label sits in a <span>), so the audit installs its own copy of axe-core in each frame it measures and takes the accessible name axe computes there. The page's own window.axe, if any, is left as it was. Installing is bounded (10 s for the main frame, 1 s per child frame). A frame that refuses the copy (a blocking CSP, a page that removes it) is measured without names. The result says so: a WARNING: accessible names could not be measured for N of these line, elementsWithoutMeasuredName in the structured summary, and elements.unmeasuredNames in the JSON report. In those frames the snapshot's own names stand, including literal slash-delimited names such as / and /docs/, so a missing-accessible-name finding for a control named only through its children may be a false positive; treat the name findings of a run with a non-zero count as partial and confirm them by hand. A frame that times out or stops answering is not evaluated by any check: its elements are counted as unresolved (with their own warning), and if that leaves nothing evaluated the audit fails rather than report a clean page. Every read of a frame is bounded by the same budgets (10 s main frame, 1 s child frame; the lookup that tells which frame owns an element gets the larger one), so a frame that stops answering cannot hold the audit open. At most four frame operations per page stay in flight across overlapping audits; if four timed-out operations are still running, no more frame work starts and the result says where measurement stopped.

Bounds: maxElements (default 400) caps how many screen-reader-reachable accessibility tree elements are analyzed. The snapshot also refs aria-hidden subtrees, which no check reports, so measuring continues past them until the budget is filled with reachable elements (up to a hard ceiling of twice maxElements measured, so a page built mostly of hidden refs stays bounded). maxFindingsPerCheck (default 20) caps the findings listed per check. Both truncations are stated in the summary and the JSON report, and the full counts are always reported. Always writes a JSON report (default filename: audit-screen-reader-{timestamp}-{token}.json).

Example flow:

1. Navigate to the target page and let it fully load
2. Run audit_screen_reader (optionally raise maxElements for a large page)
3. Fix the reported elements by ref, then re-run to confirm

Navigation Tools

browser_navigate

Navigate to a URL.

• Parameters: url (string)
• Non-2xx main-document responses are shown as an HTTP status line in page state.
• If a page opens a dialog while loading, navigation returns the pending dialog so you can resolve it with browser_handle_dialog. It does not dismiss the dialog automatically, and the interrupted response omits the page.goto replay snippet. A navigation attempt while a modal is already open fails with handling guidance. If the interrupted navigation later fails, the error appears in page console messages unless the document has changed.

browser_navigate_back

Go back to the previous page.

browser_navigation_timeout

Set default navigation timeout for existing tabs.

• Parameters: timeout (in ms; 30000-300000)

browser_default_timeout

Set default operation timeout for existing tabs.

• Parameters: timeout (in ms; 30000-300000)

Page Interaction Tools

browser_snapshot

Capture accessibility snapshot of the current page (better than screenshot for analysis). Large data: URL payloads in snapshot output are truncated to their media type prefix. AI snapshots mark a visually present subtree excluded from accessibility queries with [aria-hidden] on its boundary element. Descendants are not marked again.

• Parameters: compress (optional boolean, default false), boxes (optional boolean; overrides snapshot.boxes for this call)
  • When compress is true, repeated non-interactive ARIA snapshot nodes are collapsed in the rendered response when a repeated structural pattern appears more than 100 times. The first 10 examples of each collapsed pattern are kept.
  • Use browser_evaluate() to retrieve the full uncompressed list when needed.
  • When boxes is true, each element includes [box=x,y,width,height] in viewport-relative CSS pixels.

browser_find

Search the current page accessibility snapshot without returning the full snapshot.

• Parameters: text (case-insensitive substring) or regex (regular expression, supports /pattern/flags)
• Returns matching snapshot lines with surrounding context, shown under their path from the root of the tree; ... marks truncated off-path context.

browser_click

Perform click on a web page element.

• Parameters: element (description), ref (element reference), doubleClick (optional)

browser_type

Type text into editable element.

• Parameters: element, ref, text, submit (optional), slowly (optional)

browser_hover

Hover over element on page.

• Parameters: element, ref

browser_drag

Perform drag and drop between two elements.

• Parameters: startElement, startRef, endElement, endRef

browser_drop

Simulate an external drag and drop of files or clipboard-like data onto an element, for testing drop zones that never see a drag start inside the page.

• Parameters: element, ref, paths (optional array of absolute file paths), data (optional map of mime type to value, e.g. {"text/plain": "hello"})
• At least one of paths or data is required.
• Fails if the target's dragover handler does not accept the payload.
• paths are read from the filesystem of the machine running the server, exactly as browser_file_upload does, and a relative path resolves against the server's working directory. Unlike browser_file_upload this needs no file chooser to be open, so any page with a dragover handler is a valid target — treat it as a tool that can hand local file contents to the page.

browser_select_option

Select an option in a dropdown.

• Parameters: element, ref, values (array)

browser_fill_form

Fill multiple fields with one call.

• Parameters: fields (array of objects with name, type, ref, and value)

browser_press_key

Press a key on the keyboard.

• Parameters: key (e.g., 'ArrowLeft' or 'a')

browser_start_recording / browser_stop_recording

Record browser actions and return them as Playwright JavaScript. Start the server with --caps devtools, call browser_start_recording, perform the flow, then call browser_stop_recording.

Multi-tab recordings include the context.newPage() declarations needed by generated page aliases. Recorded assertions include the playwright/test expect setup they need to run.

Handshake-free HTTP clients must first call browser_session_open, then pass its browserSessionId to both recording tools so the recording survives across requests. Modes that cannot open separate browser sessions, such as --extension and non-isolated CDP attach, need a stateful MCP connection for recording.

browser_evaluate

Evaluate a JavaScript expression on the page, or on a specific element when a ref is provided. The function's return value is serialized back as the result.

• Parameters: function (e.g., () => document.title or (element) => element.textContent), element (optional), ref (optional)
• element and ref must be supplied together, or not at all; supplying one without the other is rejected.
• A bare expression is also accepted and is wrapped automatically: document.title behaves like () => document.title, and, when element and ref are both given, element.textContent behaves like (element) => element.textContent. The parameter is always named element.
• Whether the input is a function or an expression is decided from its source form, never from what it evaluates to, so an expression such as window.open is returned rather than called.

Screenshot & Visual Tools

browser_take_screenshot

Take a screenshot of the current page.

• Parameters: filename (optional), type (png, jpeg, or webp), scale (css or device, default css), fullPage (optional), element/ref pair (for element screenshots)
• scale: device captures a high-resolution screenshot using device pixels (accounts for the device pixel ratio); scale: css keeps the image sized in CSS pixels.
• An empty capture is an error, and its output file is removed, including automatically named files. The requested format is never silently changed. If a WebP capture is empty, reduce its dimensions or explicitly request PNG/JPEG.

browser_pdf_save

Save page as PDF.

• Parameters: filename (optional, defaults to page-{timestamp}-{token}.pdf)

This tool requires --caps pdf in the CLI.

browser_install

Install the configured browser engine (use when browser executable is missing).

• Parameters: none

Disabled by default. Enable it at server startup with --caps install, PLAYWRIGHT_MCP_CAPS=install, or "capabilities": ["install"] in the config file. Explicit core-install settings remain supported as a deprecated alias; use install in new configurations. Without this opt-in, the tool is neither listed nor callable; existing browser installations can still be used.

This tool invokes Playwright's installer, which downloads executable code. In Playwright 1.63.0, browser archives have no checksum or signature verification before extraction; the default download hosts use HTTPS. Only enable installation when you trust the download source and TLS configuration, including any custom PLAYWRIGHT_DOWNLOAD_HOST, browser-specific host overrides, or TLS-inspecting proxy. Do not disable TLS certificate validation.

For deployments that require independently verified binaries, provision the browser through your trusted deployment process and use --executable-path or an existing browser connection. The capability opt-in limits MCP-triggered installation; it does not add archive verification or change manual, CI, or Docker build downloads.

Browser Management

browser_close

Close the page.

browser_resize

Resize the browser window.

• Parameters: width, height

browser_emulate_media

Emulate CSS media features on the current page without resetting omitted features.

• Parameters: colorScheme (light or dark), reducedMotion (reduce or no-preference), forcedColors (active or none), contrast (more or no-preference), and media (screen or print); provide at least one.

Tab Management

browser_tabs

Manage browser tabs in one tool.

• Parameters: action (list, new, close, select) and optional index (for close and select).

Browser Session Tools

Following the MCP 2026-07-28 stateless prescription, browser state can be named by an explicit server-minted handle instead of living implicitly in the connection. Every browser tool except the two session tools accepts an optional browserSessionId argument; when it is omitted, the tool runs in the default session and behaves exactly as before.

browser_session_open

Opens a separate browser session — its own browser context with its own tabs, cookies and storage — and returns its opaque handle (bs_...) both in the result text and as structuredContent.browserSessionId. Pass that handle as the browserSessionId argument of other browser tools to run them in this session.

How the separate context is provided depends on the mode:

• Default persistent-profile mode: each session runs in its own fresh, disposable profile (removed when the session closes or expires); only the default session uses the stable persistent profile, whose sign-in state keeps surviving restarts. This is required — one profile directory can back only one running browser at a time.
• --isolated, remote endpoints, and CDP/--cdp-launch with --isolated: each session gets its own fresh browser context. In --cdp-launch mode each context launches its own instance of the configured application on its own free port — which is why combining --cdp-launch-port with --isolated also rejects browser_session_open: a pinned port can serve only one launched instance, so a second session would silently attach to the first session's application. A second concurrent browser context on the pinned port (e.g. from a parallel client) is likewise rejected with an error rather than attaching to the first context's application.
• Modes that reuse one live browser context — CDP attach or --cdp-launch without --isolated, --extension, the VS Code bridge, and servers created with a custom context getter — cannot create a separate context, so browser_session_open is rejected with an explanation instead of handing out a handle that would share the same tabs, cookies and storage as everything else. The same applies in the default mode when --user-data-dir pins all browsing to one user-supplied profile.

In --vscode serving, browser sessions are host-scoped: browser_session_open, browser_session_close, and every call carrying a browserSessionId always run against the default provider's session registry at the host, regardless of any browser_connect provider switch. A handle opened before a switch keeps working (and can be closed) while the proxy is switched to a VS Code-connected browser, and a session opened while switched is created by the default provider — the VS Code-connected browser itself reuses one live context and cannot host separate sessions. Only session-less tool calls follow the switch.

browser_session_close

Closes a session opened with browser_session_open and releases its browser resources.

• Parameters: browserSessionId (the handle to close)

Closing is refused with a tool error while a tool call is still running in that session — a close that disposed the browser mid-call would fail the running tool; wait for it to finish and retry.

If a pending download fails while a session closes, the response includes the download error alongside the confirmation that the session was closed. That session handle has already been removed.

Sessions that stay idle expire automatically after 30 minutes; the timer is refreshed on every use and while a tool is running in the session (overlapping calls each count, so the session survives until the last one finishes), so a long audit_site crawl is never expired mid-run. Set PLAYWRIGHT_MCP_BROWSER_SESSION_TTL_MS to override the idle TTL in milliseconds (0 or a negative value disables expiry). Passing an unknown or expired handle produces a tool error pointing back to browser_session_open; the error deliberately does not list other open sessions' handles, since handles are bearer tokens that route tool calls into their sessions. With --save-session, logged tool calls record the browserSessionId they were routed with, and recorded user actions from an explicit session carry the same browserSessionId in their logged args, so entries from different sessions stay distinguishable (default-session entries stay untagged).

Information & Monitoring Tools

browser_console_messages

Returns all console messages from the page. Large data: URL payloads in console messages are truncated to their media type prefix.

browser_network_requests

Returns all network requests since loading the page, numbered so a single one can be inspected with browser_network_request. Large data: URL payloads in request URLs are truncated to their media type prefix. When there is at least one request, a closing line points at browser_network_request.

browser_network_request

Returns credential-redacted request/response headers and body metadata for one request from the browser_network_requests listing.

• Parameters: index (the number shown in the listing, starting at 1)
• The listing is cleared by browser_navigate and when the tab closes; other navigations (link clicks, form submits, history calls) leave it in place and keep appending. Re-run browser_network_requests to get current indexes.
• Credential-bearing headers (authorization, proxy-authorization, cookie, set-cookie, x-api-key, x-auth-token) are reported as <redacted, N characters>, so their presence and size stay visible but the secret never reaches the transcript. All other headers are reported in full, one line each.
• Request and response body contents are never returned because they can contain submitted credentials or private API data. Non-empty bodies are reported as <redacted, N bytes, mime/type>; empty bodies remain <empty>.
• A request that failed after its response arrived reports both the status and the failure.
• Sections that could not be read are reported in place (<headers unavailable: ...>, <body unavailable: ...>) rather than failing the whole call; reads are bounded by the default timeout, so a still-streaming response cannot hang the tool.

Utility Tools

browser_wait_for

Wait for text to appear/disappear or time to pass.

• Parameters: time (optional), text (optional), textGone (optional)

browser_handle_dialog

Handle browser dialogs (alerts, confirms, prompts).

• Parameters: accept (boolean), promptText (optional)
• If the dialog was already closed outside the session (e.g. dismissed manually in a headed browser), the call succeeds, reports the dialog as already closed, and clears its leftover state instead of failing.

browser_file_upload

Upload files to the page.

• Parameters: paths (array of absolute file paths)
• If setFiles fails, the chooser stays available for another upload attempt; paths: [] clears the selection and completes the chooser. A successful upload clears only that chooser and waits for page activity and the configured settle delay.

browser_verify_element_visible

Verify an element by ARIA role/name.

• Parameters: role, accessibleName

browser_verify_text_visible

Verify text visibility.

• Parameters: text

browser_verify_list_visible

Verify list items at a snapshot reference.

• Parameters: element, ref, items (array)

browser_verify_value

Verify an element value or checked state.

• Parameters: type, element, ref, value

These verification tools require --caps verify:

Vision Mode Tools (Coordinate-based Interaction)

These tools require --caps vision:

browser_mouse_move_xy

Move mouse to specific coordinates.

• Parameters: element, x, y

browser_mouse_click_xy

Click at specific coordinates.

• Parameters: element, x, y, button (optional: left/right/middle), clickCount (optional), delay (optional, ms between mouse down and up)

browser_mouse_drag_xy

Drag from one coordinate to another.

• Parameters: element, startX, startY, endX, endY

Note

Coordinate-based tools require element descriptions for permission checks, but the coordinates themselves are used for action targeting.

Usage Examples

Basic Accessibility Scan

1. Navigate to example.com using browser_navigate
2. Run scan_page with violationsTag: ["wcag21aa"]

Color Contrast Check

1. Use browser_navigate to go to example.com
2. Run scan_page with violationsTag: ["cat.color"]

Multi-step Workflow

1. Navigate to example.com with browser_navigate
2. Take a browser_snapshot to see available elements
3. Click the "Sign In" button using browser_click
4. Type "user@example.com" using browser_type
5. Run scan_page on the login page
6. Take a browser_take_screenshot to capture the final state

Page Analysis

1. Navigate to example.com
2. Use browser_snapshot to capture all interactive elements
3. Review console messages with browser_console_messages
4. Check network activity with browser_network_requests

Tab Management

1. Open a new tab with `browser_tabs` and `{"action":"new"}`
2. Navigate to different pages in each tab
3. Switch to a tab with `browser_tabs` and `{"action":"select", "index": 1}`
4. List all tabs with `browser_tabs` and `{"action":"list"}`

Waiting for Dynamic Content

1. Navigate to a page
2. Use browser_wait_for to wait for specific text to appear
3. Interact with the dynamically loaded content

Note: Most interaction tools require element references from browser_snapshot. Always capture a snapshot before attempting to interact with page elements.

Development

Clone and set up the project:

git clone https://github.com/JustasMonkev/mcp-accessibility-scanner.git
cd mcp-accessibility-scanner
npm install

Playwright upgrade gate

The September 16, 2026 review keeps playwright and playwright-core paired at 1.63.0, the latest stable release on that date. Keep the local InputRecorder hub and the existing factory reference counts: multiple MCP clients share one client-side browser context, while the hub multiplexes session logs and explicit recordings and excludes sibling tool actions. A dependency bump alone must not change that ownership model.

Before adopting a stable release containing upstream #42627, adapt the hub from _enableRecorder / _disableRecorder to _startRecording({ language: 'javascript' }, sink) / _stopRecording() and verify the per-client event contract against the installed runtime. Do not ship a prerelease bump or an untested method-name fallback. Migrating to the separate connections in #42622 is a separate ownership change requiring the same lifecycle checks.

Install the pinned Chromium browser, then run the real recorder gate alongside its failure and concurrency tests:

npx playwright install chromium
npx vitest run tests/recorder.integration.test.ts tests/context.test.ts tests/browserSessions.test.ts tests/browserContextFactory.test.ts tests/tools-recorder.test.ts tests/sessionLog.test.ts

The real-browser tests cover concurrent starts, duplicate-start rejection, shared CDP clients, sibling-action attribution, stop/disconnect/restart, --save-session, and recording across stateless explicit sessions. Existing unit tests also cover failed-start recovery and overlapping start/stop. The recorder must receive the final input event before stop; unbuffered input delivered after stop begins is excluded, while buffered clicks/navigation get a 500 ms drain.

Also recheck the dependency fixes motivating the upgrade. On 1.63.0 Chromium, both full-page and oversized element screenshots changed navigator.maxTouchPoints from 1 to 0 and (pointer: coarse) from true to false; navigation restored the properties (#42617). The fixed-header/smooth-scroll retry fixture clicked successfully but emitted 19 scroll events rather than instant jumps (#42626). A candidate upgrade must preserve touch properties after full-page and element screenshots and navigation, and complete retry scrolling without smooth animation. These are dependency limitations; the recorder gate alone does not verify them.

MCP harnesses

The npm wrappers build first, then the direct harness calls every exposed MCP tool with prepared fixtures:

npm run test:mcp
npm run test:mcp:install

The Luna wrapper runs each prompt through Codex gpt-5.6-luna with xhigh reasoning. It uses a repo-scoped scanner MCP server, read-only Codex sandbox, structured PASS/FAIL evidence, and a per-prompt timeout:

npm run test:mcp:luna
npm run test:mcp:luna -- --only browser_snapshot
npm run test:mcp:luna -- --skip-optional --limit 1

Results are written under test-results/mcp-direct-harness-results/ or test-results/mcp-tool-loop-results/. Set MCP_HARNESS_RESULTS_DIR to use another output root. The Luna harness checks Codex login before starting any prompts. Its site-audit prompt uses a local HTTP fixture, not a public website. PASS requires a completed target MCP call, no failed or unfinished MCP calls, and a successful structured result. Invalid logs, failed turns, and timeouts fail the run. Non-fatal startup notices about ignored malformed local agent roles remain in the logs; they do not count as tool failures. Recovered Codex connection retries are accepted only when the turn completes successfully. Canceling the run exits with status 130.

Benchmarking tool latency

bench/mcp-bench.mjs measures what a client actually waits for: it serves a fixed synthetic site, speaks MCP to the built server over stdio, and times real tools/call round trips for navigation, interaction, snapshots and every audit tool. Build first — it runs the compiled server from lib/.

npm run build
npm run bench -- --out after.json --label after

Useful flags: --iterations <n> and --warmups <n> (defaults 5 and 1), --browser/--executable-path when the browser lives outside Playwright's own download directory, and --server <path/to/cli.js> plus --lib <path/to/lib> (supplied together) to point at a different build — that is how revisions compare:

git worktree add /tmp/baseline main && (cd /tmp/baseline && npm install && npm run build)
npm run bench -- --server /tmp/baseline/cli.js --lib /tmp/baseline/lib --out before.json --label before
npm run bench -- --out after.json --label after
npm run bench:compare -- before.json after.json

The comparison total uses only end-to-end scenarios present in both reports, so adding or removing a scenario does not distort the reported speedup.

License

MIT