k11 Selenium QA — MCP server & Claude Code plugin

Self-healing Selenium tests with an audit trail. No vendor lock-in, no rewrite to Playwright.

Most Selenium suites break the same way: someone renames an id, and dozens of find_element(By.ID, ...) calls fail overnight. AI assistants can write Selenium tests, but they guess locators, and healing tools usually patch things at runtime without telling you what changed. This MCP server gives the assistant facts instead of guesses, and gives your team a reviewable record of every fix:

• Verified locators. It opens the real page in Selenium and returns locators that match exactly one element, most stable first: data-testid, a stable id, name, link text, aria-label, then text-based XPath.
• Story → pytest. It turns "As a… I want… so that…", Gherkin scenarios or acceptance-criteria bullets into a pytest + Selenium skeleton: one test per scenario, explicit waits, and a driver fixture if you don't have one.
• Structured test runs. It runs your pytest suite and reports, for each failure, the exact locator Selenium could not find, the line that asked for it, and the page the browser was on.
• Self-healing with evidence. When the UI changes, it finds the element the test meant and suggests a verified replacement, with a confidence score, the reasons and the runner-up.
• Healing in CI. heal-ci applies only confident fixes, edits your tests and Page Objects, re-runs them, reverts anything that still fails, and opens a pull request whose description is the audit report.

Works with Claude Code (as a plugin with /story-to-test and /heal-tests), Claude Desktop, and any MCP client. Python 3.10+, Selenium 4.

See it heal in CI

The demo app's account page was redesigned: ids renamed, a "View returns" button added next to "View orders", "Edit profile" renamed to "Account settings", the cart message reworded and the page title changed. The Selenium suite uses a Page Object:

class AccountPage:
    VIEW_ORDERS = (By.ID, "view-orders")
    OPEN_CART = (By.ID, "cart-btn")

k11-selenium-qa-mcp heal-ci --apply produced this pull-request description (trimmed):

1 healed and verified, 0 suggested, 1 reverted, 1 needs review, 0 not found, 1 failure not caused by locators.

Test	Where	Old locator	New locator	Confidence	Runner-up	Re-run
test_shows_my_orders	tests/pages/account_page.py:7	(By.ID, "view-orders")	(By.ID, "orders-toggle")	82%	(By.ID, "returns-toggle") (59%)	passed

Why: name matches, role matches, text matches, type matches, tag matches; classes similar (43%).

Needs a person: "Edit profile" → best candidate "Account settings" at 45%, not confident enough to change. (By.ID, "cart-btn") healed at 85%, but the test still failed because the cart message changed too, so the change was reverted.

Not caused by locators: test_has_the_account_title (the page title changed). Untouched: it may be a real bug.

It picked "View orders" over the look-alike "View returns", fixed it where it lives (the Page Object, not the test), and refused to guess where a person should decide.

Tools

Locators are written the way your code writes them: (By.ID, "email"), driver.find_element(By.XPATH, "//button[...]"), or (By.CSS_SELECTOR, "[data-testid='save']").

Prompt: story_to_selenium, a guided workflow from story to passing test.

How self-healing works

1. Baseline. While tests pass, capture_baseline records each element's fingerprint: role, accessible name, test id, id, name, type, label, placeholder, classes, text and href.
2. Detect. run_tests loads a small pytest plugin that notes the last locator Selenium could not find, including inside WebDriverWait and expected conditions, and the page the browser was on.
3. Match. heal_locator scores every visible element on the current page against the fingerprint, with exact matching for stable attributes and fuzzy, synonym-aware matching for text and ids ("Sign in" ≈ "Log in", btn ≈ button).
4. Verify. The best match gets the most stable Selenium locator that matches exactly one element on the live page.
5. Decide. healed only when confidence is at least 0.55 and clearly ahead of the runner-up; otherwise needs_review with candidates, or not_found. Without a baseline it falls back to what the locator implies, and says so.

Commit .qa-mcp/baselines.json. Do not commit .qa-mcp/auth/ (saved logins); save_login_state adds it to .gitignore.

Pages behind a login and multi-step flows

Every page tool accepts a session:

{
  "url": "https://shop.example.com/account",
  "session": {
    "storage_state": ".qa-mcp/auth/default.json",
    "steps": [
      { "action": "click", "locator": "(By.LINK_TEXT, \"My orders\")" },
      { "action": "wait_for", "locator": "(By.CSS_SELECTOR, \"[data-testid='order-history']\")" }
    ]
  }
}

• storage_state: cookies and localStorage saved by save_login_state, in the same JSON format Playwright uses. It must be inside the project.
• steps: goto, click, fill, press, select, check, uncheck, wait_for, up to 25, with explicit waits.
• Secrets stay out of prompts and logs. Write "value": "${ENV:APP_PASSWORD}". Results show the placeholder; literal fill values are masked as •••.

Healing in CI, with an audit trail

k11-selenium-qa-mcp heal-ci --apply

1. Runs your pytest suite with the locator-recording plugin.
2. For each locator failure, opens the page the test was on (logging in first if .qa-mcp/heal.config.json says how) and heals against the committed baseline.
3. Applies a change only when confidence is at least 0.7 (--min-confidence) and clearly ahead of the runner-up. It edits every place the locator literal appears: tests, Page Objects, helpers.
4. Re-runs the affected tests and reverts any change whose test still fails, keeping the others.
5. Leaves failures that aren't about locators alone.
6. Writes heal-report.md (the pull-request body) and heal-report.json (full evidence), plus GitHub Actions outputs.

Without --apply it's a dry run. Copy docs/heal-locators.yml into .github/workflows/ to run it nightly and open the pull request, and docs/heal.config.example.json to .qa-mcp/heal.config.json.

Install

Requires Python 3.10+ and Chrome (or Edge/Firefox with QA_MCP_BROWSER=edge|firefox). Selenium Manager downloads the matching driver automatically.

Claude Code: plugin (recommended)

Adds the MCP tools plus /story-to-test and /heal-tests (needs uv):

/plugin marketplace add K11-Software-Solutions/k11techlab-selenium-qa-mcp
/plugin install selenium-qa@k11-selenium-qa

Claude Code: MCP server only

From your test project's folder:

claude mcp add selenium-qa -- uvx --from git+https://github.com/K11-Software-Solutions/k11techlab-selenium-qa-mcp k11-selenium-qa-mcp

Or with pip: pip install git+https://github.com/K11-Software-Solutions/k11techlab-selenium-qa-mcp, then claude mcp add selenium-qa -- k11-selenium-qa-mcp.

Claude Desktop and other MCP clients

{
  "mcpServers": {
    "selenium-qa": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/K11-Software-Solutions/k11techlab-selenium-qa-mcp", "k11-selenium-qa-mcp"],
      "env": { "QA_MCP_ROOT": "/path/to/your/tests/project" }
    }
  }
}

The server works on the project in QA_MCP_ROOT, or the current directory. Tests run with the project's own interpreter: QA_MCP_PYTHON, else .venv/venv in the project, else the server's Python. That environment needs pytest and selenium.

Try the demo

python demo/server.py --version v1      # http://127.0.0.1:8765/login.html (alex@example.com / correct-horse)
python demo/server.py --version v2      # the redesign

Safety

• Locators are parsed, never eval'd. Only By.<STRATEGY>, "<string>" forms (and the WebDriver strategy names) are accepted.
• Writes are confined to tests/ (write_test) or to the exact locator literal (heal-ci), which is found with Python's tokenizer, so comments and other code are never touched. Line endings are preserved.
• Tests run without a shell, with the project's interpreter.
• Healing only suggests in the MCP tools. heal-ci --apply changes code only for confident, verified fixes, and every change is in a pull request for review.

How it compares

Development

python -m venv .venv && .venv/Scripts/pip install -e ".[dev]"   # bin/pip on macOS/Linux
pytest        # unit, real-browser, MCP protocol and heal-ci tests (27)

Roadmap

• Java (JUnit 5 / TestNG) and C# (NUnit) test projects
• Healing for chained find_element calls and find_elements(...)[n]
• Visual (screenshot) fingerprints as an extra signal
• PyPI release (uvx k11-selenium-qa-mcp)

Built by K11 Software Solutions. Part of the K11 Tech Lab agentic QA series, alongside k11 Playwright QA. MIT licensed.