Top 10 Web Testing & QA MCP Servers for AI Agents (2026)

browser_snapshot(target?, filename?, depth?: number, boxes?: boolean)

Capture an accessibility snapshot of the current page, with optional depth limiting.

browser_click(element?, target, doubleClick?: boolean, button?, modifiers?: array)

Click an element referenced from the page snapshot.

browser_generate_locator(element?, target)

Generate a locator for an element to paste into a test.

browser_verify_element_visible(role, accessibleName)

Assert an element is visible, addressed by role and accessible name.

browser_verify_value(type, element, target, value)

Assert an element value. Checkboxes take the strings true or false.

Limitations

• Assertion and locator tools are absent unless you pass --caps=testing. The default tool set can drive a browser but cannot verify anything.
• Snapshot payloads dominate context. Open issues request output size limits, including #889 with 13 reactions and #806 titled Easy to reach token limit.
• A persistent browser profile can only be used by one instance at a time, so two clients sharing a workspace conflict until one passes --isolated.
• Microsoft's own README now points coding agents at the Playwright CLI instead, on the grounds that CLI calls avoid loading tool schemas and accessibility trees into context.

performance_start_trace(reload?: boolean, autoStop?: boolean)

Start a performance trace to find Core Web Vitals problems. Reloads the page by default.

lighthouse_audit(mode?)

Run a Lighthouse audit for accessibility, SEO and best practices. Mode is navigation or snapshot.

list_network_requests(pageSize?: number)

List requests made by the page since the last navigation.

take_snapshot(filePath?)

Take a text snapshot of the page built from the accessibility tree, with a uid per element.

Limitations

• Every input tool attaches a page snapshot with no way to switch it off. Issue #726, labelled confirmed, records a single click taking context from 31K to 242K tokens and a session ending at 2251% of a 200K window.
• Supports Google Chrome and Chrome for Testing only. Chrome 144 or newer is required.
• Performance tools may send trace URLs to the Google CrUX API for field data unless you pass --no-performance-crux. Usage statistics are collected by default.
• Lighthouse audits here exclude performance, which is covered by the separate tracing tools.

runPercyScan(projectName, percyRunCommand?, integrationType)

Run a Percy visual scan for a project.

fetchPercyChanges(project_name)

Fetch visual changes detected in a Percy build.

startAccessibilityScan(name, pageURL, authConfigId?: number, advancedRules?: boolean)

Start an accessibility scan and return a local CSV report path.

getFailureLogs(buildId?, sessionId?)

Retrieve failure logs for a build or session.

Limitations

• Percy tools need a Percy licence enabled by BrowserStack support, separate from a BrowserStack subscription.
• BrowserStack's docs cover VS Code, Cline, Cursor, Claude Desktop and ChatGPT. Claude Code, Windsurf, Codex and OpenCode have no documented config, and a request to publish a Claude Code plugin has been open since October 2025.
• Three separate reports of test management tools failing on accounts that require API token auth remain open on the issue tracker.
• Tool count is high enough to matter in clients that cap the tool list, such as Cursor at 40 tools.

eyes_setup_project(projectName?, apiKey?)

Set up Applitools Eyes in a Playwright JS project using the fixtures pattern.

eyes_add_checkpoints_to_test(workspaceRoot?)

Add Eyes visual checkpoints to an existing Playwright test.

eyes_get_batch_url(testOutput, workspaceRoot?)

Extract batch and session URLs from Playwright console output.

eyes_fetch_visual_results(reportPath?)

Read visual test results from the Playwright HTML report and the Eyes server.

Limitations

• The shipped package registers six tools. Applitools documentation describes a larger tool surface, so check the installed version rather than the docs page before planning around a tool name.
• Setup and checkpoint tools target the Playwright JavaScript and TypeScript fixtures SDK. Other runners are not covered.
• eyes_fetch_visual_results reads playwright-report/index.html from disk, so a run with a different reporter returns stale results from an earlier run.
• Adoption is small next to the rest of this list at roughly 754 weekly npm downloads.

analyze(url, viewportWidth?: number, viewportHeight?: number, selector?, before?: array, cookies?: array, screenshot?: object, igtTools?: array, advancedRules?)

Scan a local or remote URL with axe, optionally after interaction steps or with injected cookies.

remediate(issues: array, pageUrl?)

Return code-level fix guidance for 1 to 25 issues from an analyze response.

Limitations

• Requires a paid Axe DevTools for Web subscription and an API key. There is no free tier.
• remediate accepts 25 issues per call, so a page with hundreds of violations needs batching.
• The before parameter caps interaction steps at 20 and cookies at 20, which limits how deep into a flow a single scan can reach.
• Around 20 community axe wrappers exist, most under three stars. Check you are running Deque's image and not one of them.

list-check-stats(tag?, checkType?, search?, status?, limit?: number, page?: number, includeReliability?: boolean, range?)

List checks with reliability statistics, filtered by tag, type, status or search term.

Limitations

• Checkly documents 32 tool names but publishes parameters for list-check-stats only, so the rest have to be discovered by connecting.
• The server runs remotely and cannot read your filesystem, so writing or deploying check code stays with the Checkly CLI.
• OAuth is restricted to Checkly-approved clients. Checkly's docs name Cline, Windsurf and Devin Desktop as lacking OAuth support.
• The Model Context Protocol registry lists a third-party wrapper under an unrelated publisher rather than this server. Use the endpoint from Checkly's docs.

get-run(runId)

Fetch details for a single CI run.

get-runs(projectId, limit?: number, starting_after?)

List runs for a project, paginated up to 100 per page.

get-errors-explorer(projectId)

Group recent failures by shared error signature.

get-spec-files-performance(projectId)

Report per-spec duration and performance trends.

Limitations

• Reads results that CI already produced. It cannot run a test or open a browser.
• A read and write API key can delete runs permanently with no confirmation step. Use a read-only key unless you need writes.
• Coverage is Playwright and Cypress. Other runners need the generic reporting path.
• Useful output depends on your CI already reporting into Currents, so there is setup before the first question can be answered.

qase_get(entity, code?, id, fields?: array)

Fetch one record by type and ID. Entity is one of case, suite, run, result, plan, defect and eleven more.

qql_search(query, code?)

Search records with QQL, returning a full page in one call.

qase_case_upsert(code, id?, title?)

Create or update a test case.

qase_run_upsert(code, id?, title?)

Create or update a test run.

Limitations

• The hosted server at mcp.qase.io requires an Enterprise plan. Self-hosting over stdio works on any plan.
• Enterprise customers on dedicated infrastructure have to self-host rather than use the shared hosted endpoint.
• Published tool counts disagree between Qase's documentation and its blog. The installed package is the reliable answer.
• qase_get in a loop is roughly four times slower than one qql_search returning the same records, per the measurement in its own tool description.

Limitations

• testRigor documents its operations as prose labels such as Run green regressions rather than as tool identifiers, and publishes no input parameters. Connect the server and enumerate its tools before scripting against it.
• Hosted only. There is no local server and no package to inspect.
• No adoption signal is available: no public repo, no npm package, and no listing in the harness directories that publish install counts.
• Plain English steps move the ambiguity from the selector to the sentence, so wording changes can alter what a test does.

Limitations

• Cypress documents 13 tool names but no input parameters, so filter arguments have to be discovered by connecting.
• Read only. It reports on runs that already happened and cannot author or execute a test.
• Rate limited to 100 tool requests per hour on every plan, including paid tiers.
• Flaky test reporting, accessibility violations and UI Coverage each need their own subscription even though the server is included on all plans.