vibe-testing

imagenpm version imageNode.js imageMCP imageLicense: MIT imageCI

Code-aware browser testing for AI coding agents.

Product site · npm · MCP Registry · Launch status

vibe-testing reads your codebase so tests use your real routes and field names, runs them in a real Playwright browser, remembers what broke, and tells you what your last change fixed or regressed. It works as an MCP server that gives your editor (Claude Code, Cursor, Windsurf, VS Code Copilot, Roo Code) 14 testing tools, or as a standalone CLI.

The repository is also an Agent Plugin with a release-qa skill. Compatible agents get both the testing tools and a senior-QA workflow for coverage, evidence, regression checks, and a clear release verdict.

cd /path/to/your/project
npx vibe-testing@latest init

Then open your editor and say:

"Scan this codebase and test it against http://localhost:3000"

What code-aware testing adds

Playwright MCP gives your agent hands. vibe-testing gives it a testing workflow: code-derived scenarios, memory across runs, and a report. Two things a stateless browser tool cannot do:

1. Run it twice and it tells you what you broke. Every run writes .vibe/run-snapshot.json and diffs it against the previous run. Every scan writes .vibe/route-manifest.json and diffs your routes. The second run prints regressions and fixes instead of a wall of results:

Changes since last run
  Fixed:      /login
  Regression: /checkout
  New:        /admin/users

The same diff reaches your editor as snapshot_diff on run_full_test and run_converge, and as route_changes on scan_codebase, so the agent can flag "checkout broke after that commit" without anyone scrolling a report. Flaky routes, working selectors, and measured timeouts are also remembered between runs.

2. Zero LLM calls inside the tool. Pass/fail verification is heuristic: URL changes, toast detection, API errors. Your editor's model decides what to test; vibe-testing does the browsing and checking. No API key, no per-run cost beyond the editor subscription you already pay for.

Contents

• How it works
• MCP setup (per editor)
• MCP tools reference
• Recommended workflow
• init command
• CLI commands
• VIBE.md, project guidance
• Configuration (vibe.config.json)
• Supported frameworks
• Memory and regression detection
• FAQ

How it works

npx vibe-testing@latest init
        |
Registers 14 MCP tools in your editor
        |
You ask: "Test the checkout flow"
        |
AI calls: scan_codebase -> get_context("checkout") -> login -> explore_page -> execute_scenario -> generate_report
        |
HTML report opens in browser with screenshots of every step

No test cases to write. The AI reads your source code to understand real field names and routes, opens a browser, tests everything, and shows you what's broken.

init also:

• Detects which AI editors you have installed
• Registers vibe-test in global editor configs (~/.claude/settings.json, ~/.cursor/mcp.json, and so on) so the tools are available in every project, every session
• Creates project-level MCP configs and AI instruction files
• Auto-detects your app's URL (reads .env, vite.config, framework defaults)
• Creates VIBE.md (edit with your test credentials) and vibe.config.json
• Installs the matching Playwright Chromium build

MCP setup

Agent Plugin

Install this repository as a plugin when your agent supports the Agent Plugins standard. It includes the portable release-qa skill and starts the published npm MCP server with no API key.

• Cursor: submit or install https://github.com/AishwaryShrivastav/vibe-testing as an Agent Plugin.

• Claude Code:

  claude plugin marketplace add AishwaryShrivastav/vibe-testing
  claude plugin install vibe-testing@vibe-testing
  

• Other compatible agents: load the repository root containing plugin.json, skills/, and mcp.json.

The direct MCP and CLI setup below remains available for editors without plugin support.

MCPB bundle

Claude Desktop and other MCPB hosts can install one local bundle. Build and validate it from this repository:

npm ci
npm run build:mcpb

The artifact is written to artifacts/vibe-testing-<version>.mcpb. See the MCPB distribution guide for the bundle contents, validation steps, Smithery handoff, and the one-time Playwright Chromium prerequisite.

Option 1: automatic (recommended)

npx vibe-testing@latest init

Detects and configures all installed editors. Done.

Option 2: manual per editor

Claude Code

Add to ~/.claude/settings.json (global, works in every project):

{
  "mcpServers": {
    "vibe-test": {
      "command": "npx",
      "args": ["-y", "vibe-testing@latest", "--mcp"]
    }
  }
}

Or add to .mcp.json in your project root (project-level only):

{
  "mcpServers": {
    "vibe-test": {
      "command": "npx",
      "args": ["-y", "vibe-testing@latest", "--mcp"]
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "vibe-test": {
      "command": "npx",
      "args": ["-y", "vibe-testing@latest", "--mcp"]
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "vibe-test": {
      "command": "npx",
      "args": ["-y", "vibe-testing@latest", "--mcp"]
    }
  }
}

VS Code (GitHub Copilot)

Add to .vscode/mcp.json in your project:

{
  "servers": {
    "vibe-test": {
      "command": "npx",
      "args": ["-y", "vibe-testing@latest", "--mcp"]
    }
  }
}

Roo Code / Cline

Add to .roo/mcp.json:

{
  "mcpServers": {
    "vibe-test": {
      "command": "npx",
      "args": ["-y", "vibe-testing@latest", "--mcp"]
    }
  }
}

From local build (development)

{
  "mcpServers": {
    "vibe-test": {
      "command": "node",
      "args": ["/path/to/vibe-testing/dist/mcp-server.js"]
    }
  }
}

MCP tools reference

14 tools available to your AI editor after setup:

Tool inputs

scan_codebase

{
  "codebase_path": "/path/to/project",
  "url": "http://localhost:3000",
  "mode": "deep"
}

get_context

{ "feature": "login" }
{ "feature": "/checkout" }
{ "feature": "user profile form" }

login

{
  "email": "test@example.com",
  "password": "TestPass123!",
  "login_url": "/login"
}

scan_page_elements / explore_page

{
  "route": "/dashboard",
  "authenticated": true
}

execute_scenario

{
  "scenario": {
    "id": "create-item",
    "name": "Create a new item",
    "route": "/items",
    "steps": [
      { "action": "navigate", "url": "/items", "description": "Open items page" },
      { "action": "click", "selector": "text=Add Item", "description": "Open create form" },
      { "action": "fill", "selector": "[name='title']", "value": "Test Item", "description": "Fill title" },
      { "action": "fill", "selector": "[name='description']", "value": "Test description", "description": "Fill description" },
      { "action": "click", "selector": "button[type='submit']", "description": "Submit form" }
    ],
    "expected_outcome": "New item appears in the list",
    "requires_auth": true
  }
}

Step actions: navigate, fill, click, select, wait, assert, upload

CLI and MCP use the same scenario runner. An assert step evaluates every supplied condition:

• selector: the target must be visible. CSS, text=, label=, and placeholder= are supported; the named locators match exactly.
• value: visible text must contain this string, on the selected element or on the page body when no selector is supplied.
• url: the current URL must exactly match this absolute URL or path resolved against the configured base URL, including query and fragment.

For example:

{ "action": "assert", "selector": "#confirmation", "value": "Saved", "url": "/settings", "timeout": 5000, "description": "Check saved settings" }

These checks wait up to the step timeout (default 15000 ms per check). A failed assertion stops the scenario, records a failed step and reason, and returns status: "fail"; MCP also sets isError: true. Explicit conditions are used as the verdict after the existing authentication and form API checks. The description and expected_outcome prose are not executable assertions. Without selector, value, or url, assert only checks page health: at least 10 characters of visible body text and no nonempty visible error indicators. This compatibility check does not prove a described business outcome.

upload requires a file-input selector and a nonempty file path in value. Relative paths resolve against the codebase root. Hidden file inputs are supported. A missing file, directory, missing target, or target that is not a file input returns status: "error". The action selects one file using Playwright; add an assertion of the resulting UI to verify application-side processing. It does not automate native file-picker dialogs.

Scenarios must contain at least one step. Unknown actions are errors.

take_screenshot

{ "url": "/settings", "authenticated": true, "full_page": false }

run_full_test

{ "url": "http://localhost:3000", "codebase_path": "/path/to/project", "mode": "deep" }

run_converge

{
  "url": "http://localhost:3000",
  "max_followup_rounds": 4,
  "target_pass_rate": 0.92,
  "max_high_severity_gaps": 2
}

Recommended workflow

Full test session

Tell your AI editor:

Scan this codebase and test it against http://localhost:3000.
Log in with test@example.com / pass123. Explore the dashboard and
settings pages, run the suggested tests, and generate a report.

The AI will:

1. scan_codebase, to understand routes, forms, existing tests
2. get_context("login"), to read the actual login form source code
3. login, to authenticate in a real browser
4. explore_page("/dashboard"), clicking everything and observing what breaks
5. explore_page("/settings"), same
6. suggest_tests, to find coverage gaps
7. execute_scenario x N, running targeted test flows
8. generate_report, HTML report opens automatically
9. cleanup, closing browsers

Test a specific feature

Test the checkout flow using vibe-test. Get context for checkout,
then run the full purchase flow with card number 4242424242424242.

The AI will:

1. scan_codebase (if not already done)
2. get_context("checkout"), reading CheckoutForm.tsx, api/orders/route.ts and so on
3. login
4. execute_scenario, filling the real form fields from source code
5. generate_report

Verify a bug fix

I fixed the login redirect bug. Use vibe-test to confirm it's working.

The AI will:

1. login, testing the login flow
2. take_screenshot, visual confirmation of the post-login state
3. Report back what it sees

Find what's broken

Explore every page and tell me what's broken.

The AI will run explore_page on every route, collecting API errors, broken elements, and failed interactions, then suggest_tests with the broken items marked as high priority.

init command

npx vibe-testing@latest init [options]

What it creates:

Options:

npx vibe-testing@latest init                     # auto-detect editors, register globally + project
npx vibe-testing@latest init --no-global         # project-level only, skip global registration
npx vibe-testing@latest init --skip-browser-install
npx vibe-testing@latest init --editor cursor     # only configure Cursor
npx vibe-testing@latest init --editor claude-code windsurf

After init, edit VIBE.md with your login URL and test credentials.

CLI commands

# Set up in current project
npx vibe-testing@latest init

# Install or repair the matching browser build
npx vibe-testing@latest install-browser

# Run tests against a URL
npx vibe-testing@latest run http://localhost:3000
npx vibe-testing@latest run https://staging.myapp.com --mode deep
npx vibe-testing@latest run http://localhost:3000 --codebase /path/to/project --scope /login /dashboard

# Iterative testing until coverage thresholds
npx vibe-testing@latest converge http://localhost:3000
npx vibe-testing@latest converge http://localhost:3000 --max-rounds 6 --target-pass-rate 0.95

# Open last report in browser
npx vibe-testing@latest report

# Reset memory and screenshots for a clean run
npx vibe-testing@latest reset

run options

run and converge fail with a nonzero exit when no scenarios are discovered or an execution batch is empty. MCP run_full_test and run_converge return isError: true for these cases. Empty runs do not write a successful run snapshot or replace the prior report.

converge options

VIBE.md, project guidance

Create VIBE.md in your project root. vibe-testing reads it automatically on every run.

## Login URL
/login

## Test Credentials
- Email: test@example.com
- Password: TestPass123!

## Never Automate
- delete account
- cancel subscription
- [data-testid="danger-zone"]
- .billing-section

## Known Flaky
- /notifications (WebSocket dependent, skip or expect retry)
- /live-feed

## Notes
- Admin panel at /admin, use admin@example.com / adminpass
- Dashboard data loads async, wait for [data-loaded="true"]
- Profile page: click "Edit Profile" before form fields appear

See VIBE.example.md for the full template.

Configuration

vibe.config.json

Created automatically by init with auto-detected URL. Edit as needed:

{
  "url": "http://localhost:3000",
  "mode": "deep",
  "auth": {
    "strategy": "credentials",
    "login_url": "/login",
    "credentials": {
      "email": "test@example.com",
      "password": "TestPass123!"
    }
  },
  "never_interact": [
    "delete account",
    "cancel subscription",
    "[data-testid='danger-zone']"
  ],
  "scope": {
    "include": ["/**"],
    "exclude": ["/admin/**", "/api/**"],
    "max_routes": 30,
    "seed_routes": ["/live/dev-mode-a-now"]
  },
  "browser": {
    "headed": true,
    "slowMo": 40,
    "timeout": 30000
  }
}

Supported frameworks

Existing test files are also read to build a coverage map: Jest, Vitest, Playwright, and Cypress suites are all parsed.

Memory and regression detection

vibe-testing learns across runs and stores state in .vibe/:

• Working selectors: remembers [name='email'] worked on /login, uses it next run
• Route timings: adjusts timeouts based on measured load times
• Auth credentials: saved after first login, reused automatically
• Flaky routes: tracks high fail-rate routes, marks them for retry
• Skip routes: routes that consistently error (need URL params) are auto-skipped
• Route manifest (.vibe/route-manifest.json): every scan diffs against the previous one; new and removed routes surface as route_changes on scan_codebase results so the AI can cover them immediately
• Run snapshot (.vibe/run-snapshot.json): every run captures per-route pass/fail and diffs against the prior run; snapshot_diff flags newly_passing (fixes), newly_failing (regressions), still_failing, plus added and removed routes

{
  "snapshot_diff": {
    "newly_passing": ["/login"],
    "newly_failing": ["/checkout"],
    "still_failing": [],
    "new_routes": ["/admin/users"],
    "removed_routes": []
  }
}

run_converge returns the same shape, so iterative runs in your editor highlight what you just broke.

Reset with npx vibe-testing@latest reset to start fresh.

What a session looks like

When you ask your editor to "test the login flow", here is what it does:

User: "Test the login flow"

AI calls:
  scan_codebase({ codebase_path: ".", url: "http://localhost:3000" })
    -> Finds /login route, LoginForm component, POST /api/auth/login endpoint
    -> Returns 8 generated test scenarios

  get_context({ feature: "login" })
    -> Returns src/app/login/page.tsx (has name="email", name="password" fields)
    -> Returns src/app/api/auth/login/route.ts (POST handler, returns { token })
    -> AI now knows the real selectors: [name='email'], [name='password']

  login({ email: "test@example.com", password: "pass123" })
    -> Opens Chromium, fills the form, clicks submit
    -> Returns: { success: true, final_url: "/dashboard", tokens_found: 2 }
    -> Returns screenshot of post-login dashboard

  execute_scenario({ scenario: { name: "Login with invalid password", ... } })
    -> Returns screenshot showing error state

  generate_report()
    -> Writes .vibe/report.html, opens in browser

AI reports: "Login works. Invalid password shows an error. All 3 login scenarios passed."

FAQ

Does vibe-testing use an LLM internally? No. It uses heuristic verification (URL changes, toast detection, API errors). Your editor's model is the brain: it sees screenshots and decides what to test next. Runs have no API cost.

What's the difference between explore_page and execute_scenario? explore_page is broad: it clicks every button and input it finds and reports the results. execute_scenario is precise: you give it specific steps and it follows them exactly. Use explore_page to find what's on a page, then execute_scenario to test specific flows.

What's get_context for? It returns the actual source code for a feature, so the AI knows [name='email'] instead of guessing #email-input. Always call it before writing test steps for a specific feature.

Does it handle SPAs with client-side routing? Yes. Playwright navigates the real browser, so client-side routing (React Router, Vue Router, and the rest) works naturally.

Does it handle login / authentication? Yes. The login tool fills credentials in a real browser, captures auth tokens from localStorage/cookies, and keeps that session alive for authenticated tests. Credentials are persisted in .vibe/memory/ and reused automatically.

Will it click "Delete Account" or other destructive buttons? No. Set never_interact in vibe.config.json or VIBE.md to blocklist dangerous actions. Any button whose text or selector matches is skipped during exploration.

Can I use it without an AI editor? Yes. vibe-test run https://your-app.com runs standalone. It scans, generates scenarios, executes them, and produces an HTML report without needing an editor.

How do I test a staging environment? Set url in vibe.config.json to your staging URL, or pass it as a CLI argument: npx vibe-testing@latest run https://staging.myapp.com.

Does it work with monorepos? Yes. init detects Turborepo/pnpm/yarn workspaces and finds the frontend app automatically.

Requirements

• Node.js >= 20 (the test suite uses vitest 4.x which requires Node 20+)
• Playwright Chromium. init installs the matching build. To install or repair it directly:

  npx vibe-testing@latest install-browser
  

run and converge check for the browser before scanning your project and print the same recovery command when it is missing.

Docker

A Node 20 + Chromium image is included for environments that prefer container-based MCP servers:

docker build -t vibe-test .
# wire into your editor's MCP config:
# { "command": "docker", "args": ["run", "--rm", "-i", "vibe-test"] }

Contributing

git clone https://github.com/AishwaryShrivastav/vibe-testing.git
cd vibe-testing
npm install
npx playwright install chromium
npm run build   # tsc -> dist/
npm run dev     # run CLI without building
npm run mcp     # run MCP server without building
npm test        # unit and isolated integration tests
VIBE_REAL_BROWSER=1 npm test  # also run real Chromium CLI/MCP regressions

See CHANGELOG.md for version history. Bug reports and feature requests: GitHub issues.

License

MIT, Aishwary Shrivastav

Links

• npm: https://www.npmjs.com/package/vibe-testing
• GitHub: https://github.com/AishwaryShrivastav/vibe-testing
• Issues: https://github.com/AishwaryShrivastav/vibe-testing/issues
• Official MCP registry: listed as io.github.AishwaryShrivastav/vibe-testing at https://registry.modelcontextprotocol.io
• imagemcpindex
• Model Context Protocol
• Playwright