convez.ai
MCP ServerDesenvolvimento WebIntermediário
gsaini

Playwright MCP para Testes E2E Determinísticos

gsaini/playwright-mcp-getting-started

Validação determinística de apps React com Playwright via MCP

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server que integra Playwright com o Model Context Protocol, permitindo validação de frontend via Node.js puro — sem LLM no loop. Demonstra três estilos de assertion (snapshots, seletores CSS, page evaluate) em 9 cenários end-to-end completos.

Instalação

$npx convez-ai gsaini/playwright-mcp-getting-started

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Testes E2E automatizados sem dependência de LLM
  • Validação de fluxos de checkout e autenticação
  • Integração de Playwright como ferramenta MCP em pipelines CI/CD

Veredito

Exemplo técnico sólido mostrando que MCP é agnóstico a IA — qualquer cliente que fale o protocolo consegue dirigir um servidor Playwright. Útil para quem quer entender MCP além do Claude ou explorar testes E2E programáticos sem frameworks de teste tradicionais.

README

A tradução para o português deste README ainda está na fila.

Validating a frontend with Playwright MCP — no LLM in the loop

Node.js pnpm React Vite Tailwind Playwright MCP Biome Tests

This demo shows that the Model Context Protocol is a transport, not a runtime detail of any particular AI. Anything that speaks MCP can drive an MCP server. Here, a plain Node.js script — no Claude, no OpenAI, no inference of any kind — drives @playwright/mcp to run an end-to-end validation of Nimbus Gear, a React + Tailwind 4 demo store.

 ┌──────────────────────┐     stdio / JSON-RPC      ┌──────────────────────┐
 │  validator/run.mjs   │ ────────────────────────► │  @playwright/mcp     │
 │  (deterministic)     │ ◄──────────────────────── │  (Chromium driver)   │
 └──────────────────────┘   tools/list, tools/call  └──────────────────────┘
            │                                                  │
            │  asserts on returned text / JSON                 │  navigates,
            ▼                                                  ▼  clicks, types
   PASS / FAIL summary                              Vite dev server :5173
                                                    (React 19 + Tailwind 4)

The validator decides which tool to call next using ordinary control flow, exactly as a hand-written E2E test would. The MCP server is the only "smart" piece — it knows how to drive Chromium.

The demo app — Nimbus Gear

A small React storefront with:

  • Mock username/password auth (demo / demo) with protected routes
  • Product catalogue with search, category filter, and sort
  • Product detail pages with quantity stepper, "Add to cart", "Buy now"
  • Shopping cart with line-quantity controls, subtotals, and a Remove action
  • Multi-field checkout with inline validation
  • Order success page with generated order number
  • Light / System / Dark theme toggle (persisted to localStorage)

Built with React 19, React Router 7, Vite 8, and Tailwind CSS 4 (CSS-only theming via @theme + a dark custom variant).

Layout

PathWhat it is
app/index.html, app/vite.config.jsVite entrypoint + config
app/src/main.jsxReact root, provider tree
app/src/App.jsxRouter with protected routes
app/src/routes/6 route components (Login, Catalog, ProductDetail, Cart, Checkout, OrderSuccess)
app/src/components/Header, ThemeToggle, ProductCard, ProtectedRoute
app/src/hooks/useAuth, useCart, useTheme contexts
app/src/data/products.jsIn-memory product catalogue
app/src/styles.cssTailwind import + theme tokens (light/dark)
validator/run.mjsOrchestrator — connect, run groups, summarise
validator/lib/mcp-client.mjsWraps the official @modelcontextprotocol/sdk Client
validator/lib/snapshot.mjsParses the YAML-ish accessibility tree returned by browser_snapshot
validator/lib/helpers.mjsHigh-level helpers — clickByRole, typeSelector, evaluate, setReactInputValue, …
validator/lib/harness.mjsScenario runner + assertions (assert, assertEqual, assertContains)
validator/scenarios/auth.mjsLogin, validation, redirect
validator/scenarios/catalog.mjsSearch, filter, sort, navigation
validator/scenarios/cart.mjsAdd, quantity, remove, totals
validator/scenarios/checkout.mjsForm validation, happy path, order number
validator/scenarios/theme.mjsLight / dark / system, persistence
validator/scenarios/visual.mjsFull-page screenshots
validator/features/*.featurePlain-English specs (hand-written) — compile to .mjs via pnpm spec:compile
tools/compile-scenarios/LLM-powered .feature → .mjs compiler (Anthropic / Ollama)
biome.jsonBiome config (lint + format)

Quick start

pnpm install
pnpm exec playwright install chromium   # one-time, ~150 MB
pnpm demo                               # spawns Vite, runs all scenarios, tears down

If Vite is already running and you just want to iterate on tests:

pnpm app          # terminal 1
pnpm validate     # terminal 2

Pass --headed to watch the browser:

node validator/run.mjs --start-app --headed

Screenshots land in screenshots/ (catalogue in light + dark themes, plus a product-detail capture).

Plain-English specs (optional)

Two directories, two responsibilities:

validator/features/   intent-level .feature specs   (hand-written; source of truth)
validator/scenarios/  generated .mjs scenarios      (committed; CI runs these)

The compiler is an agentic LLM that drives the live app via Playwright MCP at compile time to discover the real DOM, then emits a deterministic .mjs. At runtime (pnpm demo) the saved .mjs runs with no LLM — that's the whole point.

COMPILE TIME (occasional)                   RUNTIME (every pnpm demo / CI run)
┌────────────┐                              ┌──────────────┐
│ .feature   │ ─────┐                       │ scenarios/   │
│ (intent)   │      ▼                       │ *.mjs        │
└────────────┘  ┌────────┐                  │ (committed)  │
                │  LLM   │                  └──────┬───────┘
                │ + tools│                         │
                └───┬────┘                         ▼
                    │ browser_navigate,    ┌──────────────────┐
                    ▼ snapshot, click…    │  validator/run   │
       ┌──────────────────────┐            │  (deterministic) │
       │  @playwright/mcp     │ ◄───────── └─────┬────────────┘
       │  ↳ Vite app on :5173 │                  │ same MCP server
       └──────────────────────┘ ◄────────────────┘ same tools
                    │
                    ▼ write_scenario(code)
              scenarios/*.mjs

Compiling

# Anthropic (default)
ANTHROPIC_API_KEY=sk-... pnpm spec:compile

# Local Ollama (no API key, no network egress; needs a tool-use-capable model)
pnpm spec:compile:ollama --model=qwen2.5-coder:14b

# Print the plan only — no Vite spawn, no model call
pnpm spec:compile:dry-run

# Subset of features
pnpm spec:compile --only=auth,catalog

# Raise the per-feature tool-use cap if a complex feature needs it
pnpm spec:compile --max-turns=80

The compiler is intentionally SDK-agnostic — it speaks raw HTTP to both providers, so the project carries no @anthropic-ai/sdk / ollama package to track or upgrade.

What the compiler does, step by step

  1. Spawns vite app (the demo app) and @playwright/mcp (the browser driver).
  2. For each .feature file, navigates the browser to a fresh state and hands the LLM a tool-use loop with these tools:
    • browser_navigate, browser_snapshot, browser_click, browser_type, browser_press_key, browser_wait_for, browser_evaluate — proxied straight through to MCP.
    • write_scenario({ code }) — terminal tool. The LLM calls this exactly once when it has explored enough to write a complete .mjs.
  3. Captures the code argument, prepends an AUTO-GENERATED header, and writes validator/scenarios/<name>.mjs.
  4. Tears down Vite + MCP.

Determinism trade-offs

  • At compile time the LLM observes a live browser. JSX with dynamic classNames, conditional rendering, computed aria-label strings — all resolved. The model writes selectors against what it saw, not what the source code says.
  • At runtime the generated .mjs is plain code. Same Playwright MCP server, same helpers, no model — every run is identical given the same app build.
  • Cost is paid once per compile and amortized over every CI run. Anthropic-side caching makes files 2–N in a batch ~10× cheaper than file 1 (look for cache hit in the per-feature log line).

Edit the .feature and recompile — do not hand-edit the generated .mjs.

Tagging scenarios

Tags use standard Gherkin @tag syntax — one or more whitespace-separated tags on the line directly above a Feature: or Scenario:. Tags on Feature: cascade to every scenario in the file; tags on Scenario: are local to that scenario.

Keep tags to two axes that map to real commands. Avoid area tags (@auth, @cart) — the filename already encodes that, and --only= covers file-level filtering.

AxisTagMeaning
Priority@smokeRuns on every PR — keep the set tiny and fast
Priority@regressionFull nightly / pre-release suite
Priority@slowExpensive scenarios (e.g. visual diffs) — skip in fast loops
Lifecycle@wipCompiler skips; not ready for CI
Lifecycle@flakyValidator soft-fails or retries; under investigation

Example:

@smoke
Feature: Authentication
  ...

  Scenario: valid credentials redirect to the catalogue
    ...

  @flaky
  Scenario: session persists across reloads
    ...

Filter at the command line — --tags= intersects with --only=:

pnpm spec:compile --tags=@smoke              # compile smoke set only
pnpm spec:compile --tags=@smoke,@regression  # union: either tag
pnpm spec:compile --tags=@smoke --tags-not=@flaky  # smoke minus flaky
pnpm validate --tags=@smoke                  # filter the run, not the compile

The compiler emits resolved tags into the generated .mjs as a tags: [...] array on each scenario(...) call, so the validator can filter at run time without re-parsing .feature files. Resolved means cascade-merged: a @smoke Feature: with a @flaky Scenario: ends up with tags: ["smoke", "flaky"].

When to skip tags entirely: if file-level (--only=auth,catalog) granularity is always enough, don't introduce tags — they're a second filtering surface that pays off only once you need sub-file control (one slow scenario inside an otherwise fast feature, a single flaky case you want to quarantine without yanking its siblings).

Lint & format

Biome handles JS / JSX / JSON; markdownlint-cli2 handles Markdown.

pnpm lint         # Biome lint
pnpm lint:fix     # Biome lint + autofix
pnpm lint:md      # markdownlint
pnpm lint:md:fix  # markdownlint + autofix
pnpm format       # Biome format (rewrites to canonical style)
pnpm check        # Biome check + markdownlint — CI-friendly, runs both gates

Rule configuration:

  • Biome: biome.json — JSX-aware, ES module style, 100-col wrap.
  • markdownlint: .markdownlint-cli2.jsonc — default rule set with MD013 (line length) and MD033 (inline HTML) disabled because the README intentionally uses wide prose and shields.io badge markup. MD024 is restricted to siblings_only so the same heading text can appear under different parents. OpenSpec-managed markdown (.claude/commands/opsx/, .claude/skills/openspec-*/, openspec/) is excluded — it has its own openspec validate.

Spec-driven changes (OpenSpec)

Larger changes are planned with OpenSpec, a spec-driven-development layer for AI coding assistants. You describe a change in plain English; OpenSpec scaffolds a proposal, design, task list, and delta specs under openspec/. You implement against the tasks, then archive — merging the deltas into the source-of-truth specs in openspec/specs/.

It's wired into Claude Code as slash commands (restart the IDE after install to load them):

CommandWhat it does
/opsx:propose "<idea>"Create a change, generate proposal + design + tasks
/opsx:applyImplement the tasks for a change
/opsx:archiveArchive a finished change, merge its delta specs
/opsx:explore, /opsx:syncBrowse changes / sync deltas into main specs

OpenSpec is a devDependency, so the CLI runs through pnpm:

pnpm exec openspec list        # active changes
pnpm exec openspec validate    # structural check of specs + changes

The generated slash commands call pnpm exec openspec for the same reason. The pnpm-workspace.yaml entry approves OpenSpec's cosmetic postinstall so pnpm 11's build-script gate doesn't block pnpm exec.

MCP-LIVE — authoring with a live browser

MCP-LIVE = author / repair compiled tests with a real browser open, observing the real DOM at each step. The mcp__playwright__* tool family drives a Playwright browser session against the running app (http://localhost:5173 locally, https://<env>.frado.ai in deployed environments), takes DOM snapshots, runs browser_evaluate to inspect specific nodes, and the resulting locators go straight into the committed test files — validator/scenarios/ *.mjs in this repo (the same role tests-compiled/**/*.spec.ts plays in TypeScript Playwright layouts).

It's not a separate runtime. The compiled tests still execute via plain Playwright in CI. MCP-LIVE only changes how the tests are written.

AUTHORING (MCP-LIVE)                       RUNTIME (CI / pnpm demo)
┌──────────────────┐                       ┌──────────────────┐
│ human or LLM     │                       │ scenarios/*.mjs  │
│ at the keyboard  │                       │  (committed)     │
└────────┬─────────┘                       └────────┬─────────┘
         │ mcp__playwright__*                       │ plain Playwright
         ▼  (snapshot, evaluate, click)             ▼  (no MCP author loop)
┌──────────────────┐                       ┌──────────────────┐
│ live browser     │                       │ headless browser │
│ on real app      │                       │ on real app      │
└──────────────────┘                       └──────────────────┘
         │
         ▼  paste locator
   scenarios/*.mjs

Two forms of the same pattern:

  • Automated — the LLM compiler drives the MCP browser, explores, and emits the .mjs. Run it when a .feature changes.
  • Manual — open the MCP browser yourself when a selector goes flaky or you're sketching a new scenario. Snapshot the page, browser_evaluate the node you care about, paste the locator into the scenario file. Same tools, same DOM truth — just no model in the loop.

In both cases the artifact that ships is a plain Playwright file; MCP is the authoring surface, not the runtime.

What the demo validates

Six feature groups, 32 scenarios total:

GroupScenariosCoverage
auth4Redirect on no-auth, form fields render, bad creds rejected, good creds redirect
catalog9Initial render, filter pill, search, empty state, sort, out-of-stock badge, deep link
cart7Quantity stepper, add toast, badge sync, multi-product cart, totals math, remove, checkout CTA
checkout5All-empty errors, email format error, valid submission, order number format, cart cleared
theme4Light removes .dark, Dark adds it, persistence, System defers to OS
visual3Light + dark catalogue screenshots, product detail screenshot

Three patterns for asserting state

The helpers in validator/lib/helpers.mjs deliberately support three interaction styles. Pick whichever fits the element at hand:

1. Snapshot tree — browser_snapshot returns an ARIA-style tree. Find a node by role + name. Resilient to layout/CSS changes.

const { nodes } = await snapshot(mcp);
findOne(nodes, "heading", "Welcome back");
await clickByRole(mcp, "button", "Sign in");

2. CSS selector via target — every interactive tool's target parameter also accepts a unique selector. Useful when an element has a stable id / data-testid but a noisy accessible name.

await clickSelector(mcp, '#filters button[aria-pressed="false"]:nth-child(2)');
await typeSelector(mcp, "#field-email", "demo@nimbus.gear");

3. Page evaluate — run JS in the page and JSON-decode the result. Best for precise, structural assertions.

const lines = await evaluate(mcp, () =>
  Array.from(document.querySelectorAll('[data-testid="cart-line"]')).map((li) => ({
    id: li.dataset.productId,
    qty: Number(li.querySelector('[data-testid="line-qty"]').textContent),
  }))
);
assertEqual(lines, [{ id: "headphones-aurora", qty: 2 }]);

A fourth helper, setReactInputValue, uses the native value setter from HTMLInputElement.prototype so programmatic value changes correctly trigger React's controlled-input tracker (a well-known React quirk that bites people trying to clear an input via el.value = "").

Why "no LLM"?

A few practical reasons to drive Playwright MCP from a non-AI client:

  • CI determinism — the same inputs always run the same scenarios. No sampling, no token budget, no "the agent decided to skip a step today."
  • Cost & speed — no inference calls. The full 32-scenario suite runs in about 70 seconds, most of which is real browser time.
  • Auditability — the test file is the spec. Reviewers see exactly what ran, in what order, with what assertions.
  • MCP server reuse — your team already runs @playwright/mcp for an AI agent? The exact same server now also powers your test suite.

LLM-driven exploration is great for finding bugs you didn't know to look for. Deterministic MCP clients are great for preventing regressions on bugs you already fixed. They are complementary, not alternatives.

Extending

  • Add a scenario: append another await scenario("…", …) block in the appropriate file under validator/scenarios/.
  • Add a feature group: create validator/scenarios/myFeature.mjs, export a myFeatureScenarios(mcp, …) function, re-export it from validator/scenarios/index.mjs, and call it from validator/run.mjs.
  • Add a helper: drop it in validator/lib/helpers.mjs. Keep it generic — anything app-specific belongs in the scenario file.
  • Validate a different app: change APP_URL (or set the env var) and rewrite the scenarios. None of the plumbing in lib/ is app-specific.
Ver o original no GitHub

Bom para

Relacionados