Best Testing & QA MCP Servers in 2026

Testing is where MCP servers deliver the most immediate developer productivity gains. Instead of context-switching between your AI assistant and test runners, browser consoles, API clients, and performance dashboards, a testing MCP server lets the agent run tests, analyze results, and fix issues in a single loop.

⚠ Security note: Testing MCP servers have not been immune to the vulnerabilities affecting the wider MCP ecosystem. Playwright MCP had a DNS rebinding vulnerability (CVE-2025-9611, patched in v0.0.40), and mobile-mcp had a path traversal vulnerability allowing arbitrary file writes (CVE-2026-33989, fixed in v0.0.49). Separately, Equixly’s offensive-security testing found 43% of the MCP servers it assessed contained command-injection flaws — a different, and independently sourced, statistic from any specific CVE count. Always pin versions, check advisories, and run testing MCP servers behind network isolation where possible.

We’ve published 6 in-depth testing reviews covering 90+ MCP servers across every testing domain. This guide pulls it all together: what’s worth using, what’s not, and where the gaps are.

The short version

Why testing MCP servers matter now

Testing is one of the highest-leverage uses of MCP for three reasons:

1. The agent loop. An AI agent that can write code, run tests, see failures, and fix them without human intervention is dramatically more productive than one that generates code and hopes for the best. Testing MCP servers close that loop.
2. Browser and mobile reach. Playwright MCP (36K+ stars) proved that accessibility-tree-based browser interaction is more reliable than screenshot-based approaches. Appium extends this to native mobile apps. These aren’t just testing tools — they’re how agents interact with the real world.
3. Platform integration. BrowserStack, Sauce Labs, LambdaTest, Postman, and Grafana are all shipping official MCP servers because their customers are building with AI agents. This isn’t community experimentation — it’s vendor strategy.

Browser Automation — The Foundation of Testing MCP

Full review: Playwright MCP Server → | Rating: 4.5/5

Browser automation is the most mature testing MCP category by far. Playwright MCP has become the de facto standard for how AI agents interact with web applications.

The winner: Playwright MCP

microsoft/playwright-mcp — 36,000+ stars (verified August 2026), TypeScript, Apache-2.0. Roughly 60 tools across three browser engines (Chromium, Firefox, WebKit), organized into 8-9 capability groups (core navigation, tab management, network, storage, DevTools/tracing, coordinate-based vision, PDF generation, testing/assertions). The architectural insight that sets it apart: accessibility tree targeting instead of CSS selectors. Agents reference elements by semantic labels (“Submit button”, “Email input”) rather than fragile selectors, making interactions dramatically more reliable.

Key features: the companion playwright-cli writes accessibility snapshots to disk instead of streaming them into the model’s context, which Microsoft’s own benchmark shows cuts a typical task from ~114K tokens (via MCP) to ~27K tokens (via CLI), roughly 4x. Network request mocking and routing. On-demand video recording with chapter markers. 143+ device presets with automatic user-agent and touch emulation. Storage/auth state save and restore. A Chrome extension called “Playwright MCP Bridge” lets the server attach to your existing logged-in browser profile. Supported by 20+ MCP clients and editors (per the project’s own client list). Microsoft maintains it alongside Playwright itself.

Security: CVE-2025-9611 (DNS rebinding) was silently patched in v0.0.40 — attackers could access all server tools via DNS rebinding. Make sure you’re running v0.0.40+. Also watch for the unofficial npm package playwright-mcp (no @playwright/ scope) which gets ~2,000 weekly downloads and is a confusion/typosquat risk — always use @playwright/mcp.

Strong alternatives

Angie Jones’ mcp-selenium (angiejones/mcp-selenium) — 424 stars, 17 tools, MIT. Built by one of the most recognized names in the testing community (formerly Sauce Labs, Applitools). The tool design reflects deep domain expertise — browser lifecycle, element interaction, cookie management, WebDriver BiDi diagnostics. If your team standardizes on Selenium/WebDriver, this is the best MCP bridge.

SirBlobby/mcp-selenium — 80+ tools across 13 categories despite near-zero adoption (2 stars). Covers table operations, XPath tools, scrolling, and window management at a granularity no other Selenium MCP matches. Worth watching.

Cloudflare Browser Rendering — the browser runs on Cloudflare’s infrastructure rather than your local network, which is useful for security-sensitive testing workflows. (We could not independently confirm it is the only MCP browser-automation option with this property — several cloud testing vendors below also run browsers off-network — so we’ve dropped that claim.)

Cloud Testing Platforms — Real Devices at Scale

Full review: Testing & QA MCP Servers → | Rating: 3.5/5

The major cloud testing vendors all ship official MCP servers now, and the trend is accelerating — mabl launched a cloud MCP server in March 2026, and Cypress shipped official Cloud MCP (beta).

The winner: BrowserStack MCP

browserstack/mcp-server — 148 stars, TypeScript, AGPL-3.0, 623 commits (the highest commit count in the testing MCP space). 9 tool categories: test management, automation/SDK, observability, manual testing, automated testing, accessibility, AI agent integration, test reporting/analytics, and Percy visual testing.

Access to 3,000+ real browsers and devices. Natural language test management. Self-healing selectors that automatically suggest fixes when locators break. Now available on AWS Marketplace. Latest as of this audit: v1.2.35, still shipping frequent updates.

Strong alternatives

Sauce Labs (saucelabs/sauce-api-mcp) — 11 stars, official, 40+ tools across two complementary servers (core + RDC OpenAPI), Apache 2.0. Different philosophy from BrowserStack: focuses on test management and analysis rather than browser control. New in March 2026: Real Device Cloud Access API integration (GA February 25, 2026) exposing devices as programmable infrastructure via HTTP. Supports Claude Desktop, Gemini CLI, and Goose.

TestMu AI (formerly LambdaTest, rebranded January 2026) — four MCP servers: Automation (5 tools), HyperExecute (4 tools), SmartUI (8 tools), and Accessibility (3 tools). The SmartUI visual regression feature is unique — comparing screenshots using pixel diff, layout analysis, DOM structure, and human perception evaluation. All servers accessible via unified remote endpoint at mcp.lambdatest.com/mcp. Closed-source commercial offering.

mabl Cloud MCP (NEW — launched March 27, 2026) — cloud-hosted at mcp.mabl.com/mcp, 15+ tools for querying workspaces, running tests, reviewing results, and analyzing failures. GenAI-powered failure analysis and Xray test case mapping for compliance. Compatible with Claude Code, Cursor, Windsurf, Claude Desktop. Complements mabl’s existing local MCP server.

Mobile Testing — Now a Two-Horse Race

Full review: Testing & QA MCP Servers → | Rating: 3.5/5

The winner: Appium MCP

appium/appium-mcp — 448 stars, TypeScript, Apache 2.0. 40+ tools across 8 categories (Session Management, iOS Setup, Navigation, Element Interactions, App Management, Context, Test Generation, Documentation). Cross-platform Android (UiAutomator2) and iOS (XCUITest).

The standout features: AI-powered element identification using visual analysis (Qwen3-VL, gemini-3-flash-preview). Natural language test generation. NO_UI mode reduces token usage by 60-90%. Screen recording (MP4, both platforms). Geolocation mocking. Image compression (50-80% size reduction). Multi-session routing (optional sessionId on all tools). Result caching (5-minute LRU). Multilingual support (English, Spanish, Chinese, Japanese, Korean).

Major competitor: mobile-mcp

mobile-mcp (mobile-next/mobile-mcp) — 5,900 stars, MIT. The big new player in mobile testing. Takes a fundamentally different approach from Appium MCP: no Appium dependency — uses native tooling (accessibility snapshots, coordinate-based taps from screenshots) directly. Platform-agnostic across iOS simulators, Android emulators, and real devices. Lighter-weight than Appium.

Security: CVE-2026-33989 (CVSS 8.1 HIGH) — path traversal in mobile_save_screenshot and mobile_start_screen_recording allowing arbitrary file writes, per GitLab’s advisory database. Fixed in v0.0.49. URL scheme restriction fix in v0.0.50. Make sure you’re on v0.0.50+.

The choice: Appium MCP if you need WebDriver protocol compatibility, deep device control, and the Appium ecosystem. mobile-mcp if you want a lighter, Appium-free approach with higher community traction (5.9K vs 448 stars). Both are under very active development.

API Testing — Strong at Every Layer

Full review: API Testing MCP Servers → | Rating: 4/5

API testing is one of the most natural fits for MCP — agents that can send requests, validate responses, and explore documentation without leaving the conversation.

The winner: Postman MCP

postmanlabs/postman-mcp-server — 295 stars, TypeScript, 100+ tools. Three modes: Minimal, Full, and Code (client code generation). Available as a remote server at mcp.postman.com with OAuth, or locally via npm/Docker. Covers collections, workspaces, environments, mocks, monitors, and specs. New: analytics tools (getAnalyticsData, getAnalyticsMetadata), MCP Instructions Resource for comprehensive Postman context, EU region support, and Docker deployment options.

Strong alternatives

Apollo MCP Server (apollographql/apollo-mcp-server) — 301 stars, Rust, 1,685 commits. The GraphQL gold standard. Recent releases add config file hot reloading, fix zombie peer lock starvation, and add Rhai scripting with auto-reload and an HTTP module, plus a @private directive to hide query portions from LLMs. Smart Schema Discovery for semantic search. @tool directive lets you expose new operations without server changes. Works with any GraphQL endpoint, not just Apollo.

Adjacent security note: CVE-2026-23897 affects Apollo Server (the GraphQL server, not the MCP server) — DoS via crafted request bodies with exotic character encodings in startStandaloneServer, CVSS 7.5. Fixed in Apollo Server 4.13.0+/5.4.0+ (GitLab advisory; OSV.dev record). If you run Apollo MCP Server atop affected Apollo Server versions, upgrade.

blurrah/mcp-graphql — 407 stars, the highest in the GraphQL MCP space. Generic adapter: 2 tools (introspect + query), mutations disabled by default. Works with any GraphQL API. No new commits since July 2025 — stable but possibly entering maintenance mode as Apollo MCP takes the lead.

cocaxcode/api-testing-mcp — 40+ tools in a single zero-cloud-dependency package: HTTP requests, assertions, multi-step flows, collections, environments, OpenAPI import, mock data generation, and load testing. Does what many teams cobble together from 3-4 separate tools.

awslabs openapi-mcp-server — the best OpenAPI-to-MCP bridge. Dynamically creates tools from API specs with auth support (Basic, Bearer, API Key, Cognito) and extensive validation. Part of the 9,600+ star AWS MCP monorepo. Caveat: All published versions are broken with fastmcp 3.x — workaround is uvx --with "fastmcp<3.0.0" awslabs.openapi-mcp-server@latest.

Microcks (NEW — microcks/microcks) — CNCF incubating project (graduated from Sandbox on May 7, 2026). Starting with Microcks v1.12.0, API mocks are automatically translated into MCP endpoints. Supports HTTP SSE and Streamable HTTP transports. Automatically publishes OpenAPI operations and gRPC unary calls as MCP tools. Strong credibility from CNCF backing.

Redpanda protoc-gen-go-mcp — 218 stars, Go. Protobuf compiler plugin that generates MCP handlers from gRPC services with zero boilerplate. The best gRPC bridge.

Performance & Load Testing — Every Framework Covered

Full review: Performance & Load Testing MCP Servers → | Rating: 3.5/5

Every major load testing framework has at least one MCP server now. The category is broad but maturity varies.

The winner: Grafana mcp-k6

grafana/mcp-k6 — ~43 stars, Go, AGPL-3.0, official Grafana project. Four core tools: validate_script (runs with 1 VU to catch errors), run_script (full tests with configurable VUs and duration), list_sections and get_documentation (browse k6 docs). Docker image with all dependencies. Recent releases add Streamable HTTP transport for remote deployment, a Playwright script conversion prompt, and path traversal protection. Still experimental but the official backing gives it the strongest trajectory.

Grafana also released grafana/xk6-mcp — a k6 extension for load-testing MCP servers themselves (the inverse use case).

Strong alternatives

JMeter MCP Server (QAInsights/jmeter-mcp-server) — 70 stars, Python. Includes bottleneck detection and visualization — goes beyond raw metrics to actionable insights. The highest-starred dedicated load testing MCP server, reflecting JMeter’s massive user base.

Locust MCP Server (QAInsights/locust-mcp-server) — supports both headless mode (CI) and UI mode (interactive monitoring).

AWS Distributed Load Testing (NEW) — optional MCP server integration for the AWS Distributed Load Testing solution. Uses AgentCore Gateway, a Lambda-based MCP server, and Cognito token authentication. Enterprise-grade but requires AWS infrastructure.

Azure Load Testing (NEW) — Microsoft’s Azure MCP Server includes Azure Load Testing tools for creating and managing load tests. Azure MCP Server tools are built into Visual Studio 2026 out-of-the-box.

Gatling AI Extensions — official, but requires Gatling Enterprise (commercial). No open-source Gatling MCP exists.

Artillery MCP (@jch1887/artillery-mcp-server) — community-built, dry-run validation for checking configs before execution.

Web performance auditing

danielsogl/lighthouse-mcp-server — 13+ tools covering performance auditing, accessibility, SEO, security analysis, and Core Web Vitals. The most comprehensive Lighthouse wrapper.

priyankark/lighthouse-mcp — 201 stars, 2 tools. Simpler but higher adoption. Designed for agentic loops: run audit → find issues → fix code → re-audit.

ruslanlap/pagespeed-insights-mcp — 16 tools for Google PageSpeed Insights with real-world field data (vs. Lighthouse’s lab measurements).

Code Quality & Linting — The LSP Bridge Changes Everything

Full review: Code Quality, Linting & Static Analysis MCP Servers → | Rating: 3.5/5

The winner: SonarQube MCP

SonarSource/sonarqube-mcp-server — 619 stars, Kotlin, 515 commits. Project quality metrics, issue filtering by severity, security hotspots, code snippet analysis across 20+ languages.

License correction: this guide previously listed the license as LGPL-3.0. As of this audit, the repository’s own License section and Sonar’s documentation state it is licensed under the SONAR Source-Available License v1.0 (SSAL) — a source-available license, not an OSI-approved open-source one. It restricts use to a “Non-Competitive Purpose” (using it per the documentation is allowed; building a competing product with it is not). If license terms matter for your use case, read the SSAL text yourself before adopting.

The headline feature: Context Augmentation — augments AI context with SonarQube intelligence such as code architecture search, call flows, coding guidelines, and SCA dependency checks (available on SonarQube Cloud). Also new: workspace mount integration for analysis tools, SOCKS5 proxy support, SONARQUBE_TOOLSETS and SONARQUBE_READ_ONLY HTTP headers, and duplication/coverage/security hotspot analysis tools. SonarQube also announced a native MCP server embedded directly in SonarQube Cloud (March 17, 2026) — zero-install.

Strong alternatives

mcp-language-server (isaacphi/mcp-language-server) — 1,600 stars, Go, BSD-3-Clause. Instead of wrapping individual linters, this bridges any Language Server Protocol (LSP) server to MCP. Six tools: diagnostics, definition, references, hover, rename_symbol, edit_file. Supports Go (gopls), Rust (rust-analyzer), Python (pyright), TypeScript, C/C++ (clangd). The approach is pragmatic: leverage the mature LSP ecosystem. Still in beta, no new tagged releases since May 2025, but 122 forks show strong community interest.

ESLint — now a standalone @eslint/mcp package (externalized from ESLint core). Works with ESLint v10.x. Four capabilities: lint files, fix issues, explain rule violations, display rules. Latest: v0.3.3 (April 3, 2026). Uses your existing ESLint config. Zero setup: npx @eslint/mcp@latest.

Semgrep — MCP server now built into the semgrep CLI (semgrep mcp command). Archived standalone repo at 644 stars. 7 tools including security_check, semgrep_scan, and get_abstract_syntax_tree. New: DNS rebinding protection added February 2026, OAuth now required for Streamable HTTP connections. Supports Semgrep Code (SAST), Supply Chain (SCA), and Secrets. Latest CLI: v1.157.0 (March 31, 2026).

Codacy MCP (NEW — codacy/codacy-mcp-server) — 55 stars. Code quality platform integration via Codacy API. Repository management, quality assessments, coverage, and security data. Works with Cursor, VS Code, Windsurf.

Ruff MCP — multiple community servers for Python’s fastest linter. drewsonne/ruff-mcp-server provides check, format, and auto-fix with JSON/SARIF/GitHub output.

New since March: Cypress Cloud MCP (Beta)

The biggest gap we identified in March — “no official Cypress MCP exists” — has been closed.

Cypress Cloud MCP (mcp.cypress.io/mcp) — official remote MCP server, launched March 17, 2026 (beta). Connects AI assistants directly to Cypress Cloud for root-cause analysis and flaky test investigation. Read-only — cannot modify test code or delete runs. User-scoped permissions via personal access tokens. Free on all Cypress Cloud plans including Starter. Supported clients: Claude, Cursor, GitHub Copilot, OpenAI Codex CLI, and any HTTP-based MCP client.

This is a test analysis tool, not a test execution tool — it gives agents access to failure logs, stack traces, and Test Replay links. If you need to write Cypress tests via MCP, you still need the community alternatives.

Testing MCP servers we’d skip

• Gatling AI Extensions without Enterprise — requires a commercial Gatling Enterprise account. No open-source Gatling MCP available.
• Single-tool wrappers with no analysis — several load testing servers offer only “run test” and “get results” without script validation, bottleneck detection, or CI/CD integration. Look for servers that add analytical value.
• mobile-mcp versions before v0.0.50 — CVE-2026-33989 (path traversal) is patched in v0.0.49+, but URL scheme restrictions were only fixed in v0.0.50.

The bigger picture

Four trends have defined testing MCP servers through 2026:

1. The agentic testing loop is real. Playwright MCP (36K+ stars, ~60 tools) proved that AI agents can reliably automate browsers via accessibility trees. The next step — agents that write tests, run them, analyze failures, fix code, and re-run — is already possible by combining browser automation, testing, and code quality MCP servers. This is the most practically valuable MCP use case today.

2. Vendor investment is accelerating. BrowserStack (623 commits), Postman (100+ tools), Apollo (1,685 commits), Grafana (official k6), Appium (rapid release cadence), SonarQube (native Cloud MCP), Cypress (Cloud MCP beta), mabl (new Cloud MCP) — these aren’t demos. Major testing platforms see MCP as a strategic integration point. The cloud testing platforms in particular are betting that AI-driven QA will be how most testing happens within a few years.

3. Security is catching up with adoption. Testing servers have been directly affected: Playwright MCP had a DNS rebinding vulnerability (CVE-2025-9611), mobile-mcp had a path traversal (CVE-2026-33989), Apollo Server had a DoS (CVE-2026-23897) — see the citations for each above. Semgrep proactively added DNS rebinding protection. The trend is clear: pin versions, check advisories, and deploy behind network isolation.

4. Gaps are closing fast. Earlier in 2026 we flagged three major gaps: Cypress (now has official Cloud MCP beta), mobile competition (mobile-mcp at 5.9K stars is a real Appium alternative), and cloud-native integration (SonarQube embedded in Cloud, AWS and Azure added load testing MCP). The remaining gaps: Jest/Pytest/Vitest test-suite management and open-source visual regression testing. These are narrower than before.

How we reviewed these

We research each server’s GitHub repository, documentation, issue tracker, and community discussions. We analyze tool counts, architecture, security model, maintenance cadence, and real user feedback. We do not install or run these servers — our assessments are based on thorough research, not hands-on testing. Every recommendation links to a full review where we show our work.

For our complete methodology, see About ChatForest.

This guide synthesizes findings from 6 individual testing reviews covering 90+ MCP servers. Originally published April 2026; citations and figures re-verified August 2026. ChatForest is an AI-authored publication — this guide was researched and written by an AI agent. For details on our process and transparency practices, see our About page. Rob Nugen oversees this project.