MK QA Master

AI 測試大師 — your AI QA loop, from analyze to advise.

English · 繁體中文

Universal MCP server for running tests across pytest / Jest / Cypress / Go, with built-in DOM analyzer, run history, and a self-improvement coach. Stable since v1.0.0 (2026-06-02) — see Stability promise below.

A Model Context Protocol server that lets Claude Desktop / Cursor / any MCP client drive your test suite end-to-end: run tests, inspect failures (screenshot + video + trace), analyze a live URL to draft test cases, and — after each run — produce a prioritized action plan telling you exactly what to fix or write next.

Full design notes: docs/framework.md.

What's in the box

• Run tests across multiple frameworks (web + mobile + API) via a single MCP surface

• Mobile via Maestro (since v0.3.0): same MCP tools, iOS Simulator / Android Emulator / real device; YAML flows; cross-platform without rewrites

• Native API testing — two runners (since v0.6.0 / v0.6.1): two peers now share the API testing slot, each fed by the artifact your team already maintains.

  • Schemathesis (QA_RUNNER=schemathesis, since v0.6.0): point at an OpenAPI 3.x / Swagger 2.0 URL or file:// schema and get property-based fuzzed tests covering status codes, response schemas, content types, and 5xx-on-fuzz violations.
  • Newman (QA_RUNNER=newman, since v0.6.1): point at an exported Postman 2.x collection (plus optional environment / globals files) and Newman replays every request, runs the embedded pm.test(...) assertions, and returns one mk-qa-master nodeid per assertion. Newman is a system prerequisite (npm install -g newman) — it's an npm package, not pip, so it doesn't ship as a Python extra.

  Both drop into the same MCP tool surface as the web / mobile runners, and both feed the same report.json / history / flake / optimizer pipeline. Existing API tests written in pytest+httpx, Jest+supertest, Cypress cy.request(), or Go net/http/httptest still ride their existing runners — no migration needed. Pact provider verification stays on the v0.7.0 conditional roadmap.

• Failure artifacts: screenshot (base64-inlined), video, Playwright trace.zip / Maestro recordings

• Run history: every run snapshotted; HTML report shows a sparkline trend

• DOM / Screen analyzer — analyze_url for web (forms / nav / dialogs / CTAs + the API endpoints the page hits) and analyze_screen for mobile (maestro hierarchy → form / cta / tab_bar modules)

• Smart test generation (generate_test): hand it an analyzer module and it writes a runnable Playwright .py or Maestro .yaml with concrete selectors, not # TODO stubs

• Auto-retry flakes — pytest side via pytest-rerunfailures; Maestro side via custom retry wrapper (no native --reruns); flaky tests surfaced separately from real failures

• Self-improvement coach (get_optimization_plan): post-run analysis across three lenses — suite quality, MCP usability, AI generation effectiveness

• JUnit XML output for CI integrations (GitHub Actions / Jenkins / GitLab)

Install

Two paths — pick the one that matches how you'll use it.

A. Run via uvx (zero install, recommended for end users)

Add mk-qa-master to your client config without installing anything globally; uv fetches and runs it in an ephemeral environment per session:

{
  "mcpServers": {
    "mk-qa-master": {
      "command": "uvx",
      "args": ["mk-qa-master"],
      "env": { "QA_RUNNER": "pytest", "QA_PROJECT_ROOT": "/path/to/your-test-project" }
    }
  }
}

That's the whole setup. First call downloads the package; subsequent calls are cached. Switching versions: uvx mk-qa-master@0.4.1 ....

B. Install into a project venv (for contributors / hacking)

pip install mk-qa-master       # or: pip install -e . from a clone
playwright install                # only if you use pytest-playwright
pip install pytest-rerunfailures  # optional, enables auto-retry

Then point your client config at the same Python interpreter:

"command": "/path/to/.venv/bin/python",
"args": ["-m", "mk_qa_master.server"]

Verify the install (v1.4+)

mk-qa-master doctor          # human-readable check report
mk-qa-master doctor --json   # for CI gates / host-LLM consumption

Walks Python version, ffmpeg + mediamtx on PATH, core deps, [edge] extras, runner registry, and MCP tool surface. Exits 0 when nothing critical is missing (warnings about unused features don't fail), 1 when mk-qa-master can't run cleanly. Run it after a fresh install or when an MCP tool returns missing_extras.

Runner-specific prerequisites

API testing (QA_RUNNER=schemathesis)

Point the runner at any OpenAPI 3.x / Swagger 2.0 schema and Schemathesis generates property-based test cases per operation — covering response schema conformance, status code conformance, content-type checks, and 5xx-on-fuzz. Results flow through the same report.json / history / flake / optimizer pipeline as your UI tests.

End-to-end walkthrough lives in docs/walkthrough-api.md; a self-contained 3-endpoint sample lives at examples/sample_api_project/.

5-line config

"env": {
  "QA_RUNNER": "schemathesis",
  "QA_OPENAPI_URL": "https://api.example.com/openapi.json"
}

Environment variables

Standard QA_TIMEOUT_SECONDS still applies (default 600s).

API testing (QA_RUNNER=newman)

Point the runner at any exported Postman 2.x collection and Newman 6.x replays every request, runs the embedded pm.test(...) assertions, and returns one mk-qa-master "test" per assertion. Results flow through the same report.json / history / flake / optimizer pipeline as the Schemathesis and UI runners.

System prerequisite: Newman ships via npm, not pip. Install once:

npm install -g newman

There's no pip install 'mk-qa-master[postman]' extra — the runner just shells out to the newman binary on PATH. If it's missing, the runner raises a clear ImportError pointing at the npm install line.

The same 3-endpoint Library API that the OpenAPI sample targets ships as a Postman collection at examples/sample_api_project/postman-collection.json — pair it with prism mock examples/sample_api_project/openapi.yaml for a fully self-contained dev loop, or point at your own staging server.

5-line config

"env": {
  "QA_RUNNER": "newman",
  "QA_POSTMAN_COLLECTION": "/absolute/path/to/your-collection.json"
}

Environment variables

Standard QA_TIMEOUT_SECONDS still applies (default 600s).

AI Visual Challenge Solver (v0.7.0)

When backend bypass isn't an option: Claude looks at the CAPTCHA, mk-qa-master does the clicks.

Supports reCAPTCHA v2 (since v0.7.0) and hCaptcha (since v0.7.1).

The first capability in the family where the AI client's vision is load-bearing, not optional. Two new MCP tools (inspect_visual_challenge + solve_visual_challenge) detect a reCAPTCHA v2 or hCaptcha image-grid challenge on the active Playwright page, screenshot it for the multimodal AI client, accept the tile-selection the AI returns, and execute the click chain. The runner is the eyes and hands; the AI client (Claude / Cursor / Gemini / GPT-4o) is the actual solver.

When to use this — Tier 1 vs Tier 3

The built-in QA knowledge layer (get_qa_context section="CAPTCHA") codifies three tiers. Reach for them in order:

Consent gate

The solver does nothing until you explicitly opt in. Two env vars drive it:

Quick start

"env": {
  "QA_RUNNER": "pytest",
  "QA_PROJECT_ROOT": "/path/to/project",
  "QA_VISUAL_CHALLENGE_CONSENT": "true",
  "QA_VISUAL_CHALLENGE_AUTHORIZED_DOMAINS": "client-staging.example.com"
}

Then, when a run_tests call surfaces an external_dependency failure that points at a CAPTCHA, the AI client can escalate:

mk-qa-master.inspect_visual_challenge()  # screenshot + tile grid
→ AI vision picks tiles [0, 4, 7]
mk-qa-master.solve_visual_challenge(
    challenge_id="...", selected_tile_indices=[0, 4, 7], confirm=true,
)
→ status: "passed", token: "...", hint: "CAPTCHA verified. Resume your test."

Full walkthrough lives in docs/walkthrough-visual-challenge.md. PRD: docs/prd-v0.7-visual-challenge.md.

Hard-stop domains

Regardless of consent or allowlist, the solver refuses to operate on known third-party identity providers (accounts.google.com, login.microsoftonline.com, id.apple.com, facebook.com, login.live.com, etc.). No legitimate QA scenario justifies a CAPTCHA solver against someone else's login portal.

Privacy

No screenshot retention beyond the active inspect→solve cycle. Telemetry logs the boolean outcome only — never the screenshot, never the challenge text, never the tile selection. The 5-minute LRU cache holds at most 10 outstanding challenges per process and never touches disk.

Success rate caveat

The AI client's vision model does the actual judging — Claude Sonnet 4, GPT-4o, and Gemini 2.5 all ship with native vision but their accuracy on a 3x3 reCAPTCHA varies. Plan for at least one retry per challenge (reCAPTCHA gives you three before locking out). get_telemetry will eventually surface aggregate pass-rate so you can size that expectation per-client.

Scope: reCAPTCHA v2 image-grid only in v0.7.0. hCaptcha lands in v0.7.1. reCAPTCHA v3 / Cloudflare Turnstile are permanently out of scope — they don't surface a visible challenge to inspect.

OWASP API Security scanning (v0.8.0)

Schemathesis catches correctness drift. v0.8.0 adds the layer that catches the security drift hiding behind a passing schema.

v0.8.0 ships an OWASP API Security Top 10 (2023) rule-based scanner as a new MCP tool: run_api_security_scan. It loads an OpenAPI 3.x spec, walks each (path × method), and dispatches five purely-HTTP- observable rules:

API4 (rate limit DoS risk), API6 (business flow modeling), API7 (SSRF callback infra), API9 (prod recon), API10 (upstream APIs) are deferred — see docs/prd-v0.8-api-security.md §3.

Consent + authorization gates

Mirrors the v0.7 visual-challenge consent model:

The mass_assignment rule mutates server state — it's excluded from default categories. Callers must opt in: categories=["headers", "broken_auth", "bola", "function_authz", "mass_assignment"].

Quick start

"env": {
  "QA_RUNNER": "pytest",
  "QA_PROJECT_ROOT": "/path/to/project",
  "QA_API_SECURITY_CONSENT": "true",
  "QA_API_SECURITY_AUTHORIZED_DOMAINS": "api.staging.example.com"
}

Then ask the AI client to scan:

mk-qa-master.run_api_security_scan(
    spec_url="https://api.staging.example.com/openapi.yaml",
    auth={
        "token": "alice's bearer token",
        "alt_user_token": "bob's bearer token",
        "bola_test_ids": {"user_a": [101, 103], "user_b": [202]}
    },
    severity_threshold="medium"
)

Returns the v0.8 security report block:

{
  "scan_id": "a3f8d1c9b7e2",
  "spec_url": "...",
  "base_url": "https://api.staging.example.com",
  "categories_run": ["headers", "broken_auth", "bola", "function_authz"],
  "rules_ran": ["OWASP-API8-Headers", "OWASP-API2-BrokenAuth", ...],
  "ops_scanned": 23,
  "severity_threshold": "medium",
  "findings": [
    {
      "rule_id": "OWASP-API1-BOLA-CrossUserDataExposure",
      "severity": "critical",
      "endpoint": "GET /orders/{id}",
      "title": "user_a can read user_b's object id=202 — missing object-level authorization check",
      "evidence": {"actor": "user_a", "target_owner": "user_b", "target_id": 202, "probed_path": "/orders/202", "status_code": 200, ...},
      "remediation_hint": "Compare the caller's identity to the object's owner before returning..."
    },
    ...
  ],
  "summary": {"total": 7, "by_severity": {"critical": 2, "high": 4, "medium": 1, "low": 0, "info": 0}}
}

The Tier 1 ground truth

examples/sample_vulnerable_api/ ships a deliberately-vulnerable Flask app where every in-scope OWASP category has a vuln/safe endpoint pair. Run it locally to see what each rule looks like in action:

cd examples/sample_vulnerable_api
pip install -r requirements.txt
python app.py  # binds 127.0.0.1:5099
# Then from another shell, point run_api_security_scan at
# http://127.0.0.1:5099 + the bundled openapi.yaml

The scanner finds all 5 categories on /vuln/* and produces zero false positives on /safe/*. That property is enforced by the Tier 1 dogfood tests on every PR.

Security note

The scanner runs adversarial test cases. Do not point it at production systems you don't own, and do not point it at any system where you don't have authorization. The two env vars above are the contract.

PRD: docs/prd-v0.8-api-security.md. The earlier v0.8 mobile attempt was parked — see docs/v0.8-mobile-postmortem.md for what we learned and how it shaped the API-security PRD's testing gates.

Use as a Claude Code / Codex / Hermes / OpenClaw skill (v0.9.0)

Same skill folder loads in four different agent hosts via the agentskills.io convention.

v0.9.0 packages mk-qa-master as a cross-host agent skill in addition to its MCP-server form. The skills/mk-qa-master/ folder is the single source of truth — the same SKILL.md, slash commands, and reference docs load into:

• Claude Code — via .claude-plugin/plugin.json (this repo is a plugin marketplace).
• OpenAI Codex — via .codex-plugin/plugin.json (Codex reads Claude- style marketplaces).
• OpenClaw — install from local checkout: openclaw plugins install /path/to/mk-qa-master.
• Hermes Agent — symlink the skill folder into ~/.hermes/skills/.

Quick install (Claude Code)

# Inside Claude Code:
/plugin marketplace add kao273183/mk-qa-master
/plugin install mk-qa-master@mk-qa-master

Restart Claude Code so the skill registers. Then any QA testing prompt auto-activates the skill — or explicitly invoke a slash command:

/mk-qa-master:run-tests login
/mk-qa-master:generate https://staging.example.com
/mk-qa-master:api-security https://api.staging.example.com/openapi.yaml

What the skill does

The skill is a single-file operating contract that teaches the host how to drive mk-qa-master's 22 MCP tools coherently. It encodes:

• When to auto-activate — phrases like "run my tests", "why did this test fail", "scan this API for OWASP issues" trigger it.
• Five flows — run tests / generate tests / debug failures / solve CAPTCHAs / scan APIs.
• Hard rules — surface consent errors verbatim, don't silently re-run with relaxed filters, confirm before destructive runs.

Full reference at skills/mk-qa-master/SKILL.md.

Why a skill on top of an MCP server?

The MCP server makes the 22 tools callable by any client. The skill makes them discoverable + governed: it gives the host's skill router enough context to decide when to use the tools and which flow to follow. Inspired by microsoft/Webwright, which uses the same pattern.

Stability promise (v1.0.0)

22 tools. Frozen schema. Versioned drift. Pin and go.

mk-qa-master shipped v1.0.0 on 2026-06-02. The MCP tool surface is locked: 22 tools, the consent gate env vars, the plan / bookend shapes, and the hard-stop blacklists don't change without a deprecation cycle.

What this means for callers

Breaking changes require a v2.0 bump. Deprecations get ≥ 1 minor of warning with DeprecationWarning raised at runtime, "Deprecated:" in the MCP tool description, and an entry in docs/MIGRATION-1.x-to-2.0.md (created when v2.0 work opens).

How the promise is enforced

A CI snapshot test (tests/test_v1_schema_snapshot.py) freezes the 22-tool surface in tests/snapshots/v1/tool_surface.json. Any drift fails CI unless the PR sets BREAKING_CHANGE_ACK=true AND both docs/MIGRATION-0.x-to-1.0.md and docs/DEPRECATION-POLICY.md exist. The ack alone isn't a free pass — the docs must be in place.

A second test (tests/test_v1_doc_sync.py) scans every public doc for tool-count claims and fails if any disagree with the live server.

Read the contract

• docs/MIGRATION-0.x-to-1.0.md — every additive shape change v0.7 → v1.0 enumerated. TL;DR: v0.10 → v1.0 is a no-op.
• docs/DEPRECATION-POLICY.md — formal cycle. patch = bugfix, minor = additive, major = removal (only after deprecation).
• docs/prd-v1.0-stability-lock.md — locked PRD.

License Evolution Plan (v1.2.1 announcement)

MIT today. Apache 2.0 in v2.0.

mk-qa-master is announcing that it will relicense from MIT to Apache 2.0 in v2.0.0. This patch (v1.2.1) is the formal announcement and starts the deprecation clock.

What changes for you

Apache 2.0 grants strictly more rights than MIT (explicit patent grant + trademark protection) while keeping the same commercial-use permission. No scenario reduces your usage rights.

Timeline

• v1.2.1 (this release): announcement only. No code changes.
• v1.3.x onwards: still MIT. Hold cycle for at least one minor before v2.0 lands.
• v2.0.0 (TBD): actual relicense. Apache 2.0 LICENSE file, NOTICE file, source-header sweep, manifest sync.

Plus a commitment to maintain v1.x bugfix releases for ≥ 6 months after v2.0.0 ships. If your company can't move to Apache 2.0 immediately, you have a runway.

Why

Long-term sustainability — patent peace, trademark protection, contributor IP unambiguity, broader corporate procurement compatibility. See docs/RELICENSING.md for the full rationale + mechanical v2.0 checklist.

Edge AI Runner (v1.1.0+)

RTSP stream + YOLO inference + pytest assertions in a single QA_RUNNER=edge flag.

v1.1.0 adds an Edge AI Inference Runner that drops into the same analyze → generate → run loop the web and mobile runners already use. The new analyze_stream MCP tool (tool #22) probes RTSP geometry and emits candidate test cases per detected label.

Quick install

pip install "mk-qa-master[edge]"   # opencv-python + ultralytics + requests

# Plus the binary deps the runner shells out to:
brew install ffmpeg mediamtx       # macOS
# or: sudo apt install ffmpeg + download mediamtx from https://github.com/bluenviron/mediamtx

End-to-end walkthrough

The bundled sample fixture at examples/sample_edge_fixture/ exercises the full loop. Tested against mk-qa-master==1.1.0 (Edge AI), mk-qa-master==1.1.1 (housekeeping), and 1.1.2 (this doc patch).

1. Configure the runner. Three env vars are enough for the desktop path:

export QA_RUNNER=edge
export QA_RTSP_SOURCE="$(pwd)/examples/sample_edge_fixture/factory.mp4"
export QA_MODEL_PATH=yolov8n.pt    # ultralytics auto-downloads on first use

Optional tuning (defaults in parentheses): QA_MIN_FPS (25), QA_LATENCY_SLA_MS (40), QA_IOU_THRESHOLD (0.5).

2. Ask Claude / Cursor / any MCP host. With mk-qa-master wired as an MCP server (see Wire into Claude Desktop), prompt:

"analyze the stream at examples/sample_edge_fixture/factory.mp4 with the bundled annotations sidecar, then generate detection tests for each label."

Claude calls analyze_stream → gets {width: 320, height: 240, fps: 5, labels: ["forklift", "person"], candidate_tcs: [...]} → calls generate_test per label → writes test_edge_factory_person.py and test_edge_factory_forklift.py to PROJECT_ROOT/tests/.

3. Run. The runner brings up local mediamtx + ffmpeg (the file source loops over RTSP), exports EDGE_* env vars from the QA_* you set, and invokes pytest. Each generated test:

• Reads frames via cv2.VideoCapture(EDGE_RTSP_URL)
• Pushes each frame through the YOLO backend
• Tracks per-frame latency in a LatencyTracker
• Asserts the per-label detection appears within the IoU threshold for at least one frame in the ground-truth window
• Asserts p95 latency ≤ EDGE_LATENCY_SLA_MS
• Asserts sustained throughput ≥ EDGE_MIN_FPS over a 150-frame window

The report lands in PROJECT_ROOT/report.json + junit.xml, gets archived under test-results/history/, and triggers get_optimization_plan like any other runner.

Vendor-host safety default

analyze_stream refuses RTSP URLs at known surveillance / IoT camera vendor domains (Dahua, Hikvision, Ezviz, Axis, Amcrest, Lorex, Swann, Reolink) by default. Keeps accidental probing of public camera feeds off the default path. Set QA_EDGE_ALLOW_VENDOR_HOSTS=true to opt in for own-camera testing.

Resilience injection (v1.3.0)

v1.3.0 adds an opt-in network degradation harness for Edge runs. Pass resilience_mode="netem" to generate_test and the emitted pytest uses Linux tc qdisc (via mk_qa_master.edge.resilience.apply_netem) to inject 200 ms latency + 5 % packet loss on the loopback interface, asserts the runner stays within SLA under degradation, then clears the qdisc on teardown.

Double-guarded for safety:

• apply_netem raises RuntimeError on non-Linux (macOS / Windows hosts → tests pytest.skip automatically).
• Even on Linux it refuses to run until QA_EDGE_NETEM_ENABLED=true — explicit consent for the loopback impact.

The same module also ships three companion helpers: clear_netem (idempotent teardown), kill_ffmpeg_subprocess (process-loss scenario), and build_corrupted_gop_fixture (ffmpeg-driven bitstream noise injection). See src/mk_qa_master/edge/resilience.py and the v1.3.0 PRD for the full menu.

When tests run under resilience mode, the emitted report carries an additive per-test edge_metrics block (frame drops, recovery time, etc.). get_optimization_plan reads it to surface 4 Edge-specific flake signals (corrupted-frame rate, recovery-time skew, drop bursts, sustained latency violations) alongside its usual signal mix.

Phase status

Troubleshooting

Migration from v1.0.0 → v1.1.x

v1.0.0 → v1.1.0 is additive only — no existing tool changed shape. v1.1.0 → v1.1.1 → v1.1.2 are patch releases (housekeeping + docs). v1.2.0 added Phase 3 (remote inference). v1.3.0 added Phase 4 (resilience injection + Edge flake signals). See docs/MIGRATION-1.x.md for the full change log + the list of new QA_* env vars (QA_EDGE_NETEM_ENABLED, …).

Full PRD: docs/prd-v1.1-edge-ai-runner.md.

Universal plan + verify bookend (v0.10.0)

Declare success up front, run the work, get a checklist back — on every meaningful tool, not just one.

v0.10.0 generalizes the v0.9.4 bookend pattern (which lived only on run_api_security_scan) to 5 core tools. Each one accepts an optional plan_id kwarg returned by qa_plan. When you thread that plan_id through, the tool's response gains a plan_verification envelope that auto-verifies the work against the critical points you declared — no separate verify_plan call needed.

plan = qa_plan(
    task="Smoke the signup flow",
    critical_points=[
        {"id": "CP1", "verification_hint": "test_happy_path passes"},
        {"id": "CP2", "verification_hint": "BOLA-on-orders"},
    ],
)

result = run_tests(plan_id=plan["plan_id"])

# result["plan_verification"]["status"] == "passed" | "incomplete" | "failed"
# result["plan_verification"]["checklist"] tells you per-CP outcomes

Backward compat: omitting plan_id keeps the v0.9.x response shape intact. See docs/prd-v0.10-universal-bookend.md for per-tool evidence contracts and the locked decisions.

Wire into Claude Desktop (legacy MCP-only path)

If you prefer the bare MCP-server wiring (no plugin/skill layer), copy examples/configs/claude_desktop_config.example.json to:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Two environment variables drive the runtime:

Per-runner snippet

pytest-playwright:

"env": { "QA_RUNNER": "pytest", "QA_PROJECT_ROOT": "/path/to/python-project" }

Jest:

"env": { "QA_RUNNER": "jest", "QA_PROJECT_ROOT": "/path/to/node-project" }

Cypress:

"env": { "QA_RUNNER": "cypress", "QA_PROJECT_ROOT": "/path/to/cypress-project" }

Go test:

"env": { "QA_RUNNER": "go", "QA_PROJECT_ROOT": "/path/to/go-project" }

Maestro (mobile, since v0.3.0):

"env": {
  "QA_RUNNER": "maestro",
  "QA_PROJECT_ROOT": "/path/to/maestro-flows",
  "QA_ANDROID_HOST": "127.0.0.1:5555"
}

QA_ANDROID_HOST is optional — only set it when targeting BlueStacks / Genymotion / cloud-Android-farm via remote ADB. iOS Simulator / Android Emulator / local USB device auto-discovered.

Schemathesis (API):

"env": {
  "QA_RUNNER": "schemathesis",
  "QA_OPENAPI_URL": "https://api.example.com/openapi.json"
}

Newman (Postman):

"env": {
  "QA_RUNNER": "newman",
  "QA_POSTMAN_COLLECTION": "/absolute/path/to/collection.json"
}

Edge AI (RTSP + YOLO, since v1.1.0):

"env": {
  "QA_RUNNER": "edge",
  "QA_RTSP_SOURCE": "/absolute/path/to/factory.mp4",
  "QA_MODEL_PATH": "yolov8n.pt"
}

Requires pip install "mk-qa-master[edge]" + ffmpeg + mediamtx on PATH. See the Edge AI Runner walkthrough for the full env-var table and troubleshooting.

Other MCP clients

MCP is an open protocol — this server isn't Claude-only. The same Python process talks to any MCP client over JSON-RPC stdio. What differs across clients is (1) the config file format and (2) how reliably the underlying model auto-chains tool calls.

Example configs ship in the repo: codex-config.example.toml · gemini-config.example.json · claude_desktop_config.example.json.

Codex (TOML):

[mcp_servers.mk-qa-master]
command = "/path/to/.venv/bin/python"
args = ["-m", "mk_qa_master.server"]
cwd = "/path/to/mk-qa-master"
[mcp_servers.mk-qa-master.env]
QA_RUNNER = "pytest"
QA_PROJECT_ROOT = "/path/to/your-test-project"

Gemini (JSON, same shape as Claude Desktop):

{
  "mcpServers": {
    "mk-qa-master": {
      "command": "/path/to/.venv/bin/python",
      "args": ["-m", "mk_qa_master.server"],
      "cwd": "/path/to/mk-qa-master",
      "env": {
        "QA_RUNNER": "pytest",
        "QA_PROJECT_ROOT": "/path/to/your-test-project"
      }
    }
  }
}

Tool descriptions already nudge the recommended chains (analyze_url → generate_test, get_qa_context before generating domain tests). Clients with weaker tool-selection benefit most from explicit prompts that name the steps.

Tool surface

Shared across all runners (some tools degrade gracefully on non-pytest runners):

Resources

Self-improvement loop

After every run, _archive_report() snapshots report.json into test-results/history/ and writes a fresh optimization-plan.md covering:

1. Suite quality — outcomes string per test (PFPFP); transitions → flake score; 3+ identical-signature fails → broken; rerun-passed → flaky-in-run
2. MCP usability — top tools, error rates, repeat-arg patterns, common A→B chains (from telemetry JSONL logs)
3. AI strategy — adoption rate of generate_test outputs, coverage gaps from analyze_url modules with no matching test files

The plan emits prioritized actions (high / medium / low) each with target + evidence + suggestion + optional auto_action_hint the MCP client can chain into the next tool call.

Project layout

mk-qa-master/
├── pyproject.toml
├── src/mk_qa_master/
│   ├── server.py            # MCP entry (tool routing + telemetry wrap)
│   ├── config.py            # Paths + env vars
│   ├── runners/             # Per-framework plugins
│   │   ├── base.py          # TestRunner abstract interface
│   │   ├── pytest_playwright.py
│   │   ├── jest.py
│   │   ├── cypress.py
│   │   └── go_test.py
│   ├── reporters/
│   │   └── html.py          # Self-contained HTML render
│   └── tools/               # Thin shims + analyzer + optimizer + telemetry
└── tests_project/           # Example project under test

Adding a runner

1. Create src/mk_qa_master/runners/your_runner.py, subclass TestRunner, implement the abstract methods
2. Register the name in runners/__init__.py's REGISTRY
3. Done

End-to-end workflow

The intended pipeline — from a URL to "what should I improve next time":

flowchart LR
    URL[URL] -->|analyze_url| MOD[modules<br/>+ candidate TCs<br/>+ API endpoints]
    MOD -->|generate_test<br/>module=...| TEST[tests/test_*.py<br/>runnable skeleton]
    TEST -->|run_tests| RES[report.json<br/>+ screenshots<br/>+ trace.zip<br/>+ junit.xml]
    RES -->|auto archive| HIST[history/ snapshot]
    RES -->|generate_html_report| HTML[HTML report<br/>self-contained]
    HIST -->|auto write| PLAN[optimization-plan.md]
    PLAN -.->|next session reads| URL

The loop is the point: every run feeds the optimizer, the optimizer points at the weakest link, the next run hits that link first.

Walkthrough — testing a login page

In a Claude / Cursor session:

You: 分析 https://shop.example/login，幫我寫對應測試

Claude: [analyze_url] Found 1 form (email_password_form_0) + 3 API endpoints. 5 candidate TCs. [generate_test with the form module] Wrote tests/test_login.py — runnable with concrete selectors, no # TODO stubs.

You: 跑

Claude: [run_tests] 23 passed, 0 failed in 31s. Screenshots + step traces captured for every test.

You: 下一步該做什麼？

Claude: [opens report://optimization] Top: tests/test_login.py::test_invalid_credentials is flaky (flake_score=0.4, outcomes=PFPFP). Suggestion: add wait_for_response('/api/login') before asserting the error message.

The three optimizer lenses (suite quality / MCP usability / AI generation effectiveness) make every "下一步" answer data-driven, not gut feel.

Walkthrough — testing a mobile app (Maestro)

Same shape, different runner. Requires Maestro CLI installed + simulator/emulator booted + your app launched (or pass launch_app=true with app_id):

You: 分析 your mobile app 首頁的條碼按鈕、寫對應測試

Claude: [analyze_screen(app_id="com.example.app", launch_app=true)] Found 15 interactive elements; matched barcode_button (text="條碼", resource_id="barcodeButton") + candidate TCs. [generate_test with the cta module] Wrote maestro-flows/test_barcode.yaml — tapOn: { id: barcodeButton } + waitForAnimationToEnd + takeScreenshot, ready to maestro test.

You: 跑

Claude: [run_tests] 5 flows pass, retry didn't fire. Screenshots embedded in HTML report.

You: 上面這個按鈕有時候會 fail、為什麼？

Claude: [get_optimization_plan] barcode_button::barcode_button flagged flaky (flake_score=0.4, outcomes=PFPFP, rerun_count=1). Suggestion: 加 waitForAnimationToEnd 或 extendedWaitUntil 等動畫穩定後再 tap。

Mobile-specific notes:

• The same qa-knowledge.md (built-in methodology + your domain) feeds both web and mobile runs — write your business rules once.
• analyze_screen filters out iOS status bar (signal / wifi / battery) and asset-name labels (bg_*, *_filled); the result is signal-heavy.
• Maestro's takeScreenshot: <name> directive controls which screens show up as inline images in the HTML report.

Prompting cookbook

Each row shows a phrase you can paste into a Claude / Cursor session and the underlying MCP tool call it should trigger. Use as a reference for "how do I get the AI to do X without naming the tool myself."

One-time setup

Day-to-day testing

Building tests from a URL (web)

Building tests from a mobile screen (Maestro)

Requires QA_RUNNER=maestro, Maestro CLI, and a booted simulator/emulator/device.

BlueStacks / remote Android instances: set QA_ANDROID_HOST=127.0.0.1:5555 (or whatever host:port BlueStacks exposes — see Settings → Advanced → Android Debug Bridge). The Maestro runner will adb connect before each test and analyze_screen, and bumps the hierarchy timeout to 60s to absorb the slower TCP-ADB path. Genymotion / Nox / LDPlayer / WSA work the same way; any host:port that responds to adb connect is fine.

Continuous improvement

Tips — getting Claude to pick the right tool

• Mention QA knowledge explicitly — "reference qa knowledge when testing coupon" pushes Claude to call get_qa_context first; saying just "test coupon" may skip it.
• State the order — "analyze first, then write" forces analyze_url before generate_test; "just write a test for X" skips analysis.
• Batch vs precise — "auto-generate the whole page" → auto_generate_tests; "write one test per candidate_tc" → manual chain.
• Failure debugging — Asking "why did it fail / show me the screenshot" reliably triggers get_failure_details (which now returns screenshot + trace + video paths).

Anti-patterns

• ❌ "Run it 5 times to see if it's flaky" — the runner has auto-retry + history; just ask "is it flaky" and let get_optimization_plan answer.
• ❌ "Generate 100 tests" — noise > signal. Use get_optimization_plan first to find what's missing.
• ❌ "Test all edge cases" — too vague. Phrase as "test every candidate_tc for this form" — concrete, bounded, traceable.

Sample outputs

analyze_url (excerpt)

{
  "url": "https://shop.example/login",
  "page_title": "Login",
  "module_count": 3,
  "modules": [
    {
      "kind": "form",
      "name": "email_password_form_0",
      "selectors": {
        "container": "#login",
        "fields": [
          {"label": "Email", "selector": "#email", "type": "email", "required": true},
          {"label": "Password", "selector": "#password", "type": "password", "required": true}
        ],
        "submit": "button[type='submit']"
      },
      "candidate_tcs": [
        "所有必填欄位為空時送出，應顯示必填錯誤",
        "Email 欄位填入格式錯誤的字串（無 @），應顯示格式錯誤",
        "Password 欄位輸入後應預設遮蔽（type=password）",
        "全部填入合法值後送出，應觸發成功流程"
      ]
    }
  ],
  "api_endpoints": [
    {
      "method": "POST",
      "path": "/api/login",
      "status": 401,
      "candidate_tcs": [
        "POST /api/login payload 缺必填欄位應回 400 + 欄位錯誤訊息",
        "POST /api/login 合法 payload 應回 2xx",
        "POST /api/login 缺少 auth header 應回 401/403"
      ]
    }
  ]
}

generate_test output (smart, with module)

"""
Login happy path

Auto-generated from analyze_url module: email_password_form_0 (kind=form)
"""
from playwright.sync_api import Page, expect


def test_login(page: Page):
    page.goto('https://shop.example/login')
    page.locator('#email').fill('test@example.com')
    page.locator('#password').fill('TestPass123!')
    page.locator("button[type='submit']").click()
    # TC: Email 欄位填入格式錯誤的字串（無 @），應顯示格式錯誤
    # TC: Password 欄位輸入後應預設遮蔽
    # TC: 正確 Email + 正確密碼 → 導向 dashboard
    # TODO: 補上實際斷言，例如：
    # expect(page).to_have_url(...)
    # expect(page.get_by_text("成功")).to_be_visible()

optimization-plan.md (excerpt)

# Optimization Plan — 2026-05-12T14:03:40

_Based on 6 archived runs._

## Prioritized Actions

### 1. 🔴 HIGH — flaky
- **Target**: `tests/test_login.py::test_invalid_credentials`
- **Evidence**: flake_score=0.4, outcomes=PFPFP, rerun_count=1
- **Suggestion**: 加 explicit wait（wait_for_response / locator wait）

### 2. 🟡 MEDIUM — coverage_gap
- **Target**: `register_form`
- **Evidence**: 由 analyze_url 偵測但 repo 內找不到對應 test_*.py
- **Suggestion**: `call generate_test(description="...", filename="test_register_form.py")`

HTML report

Open the live rendered demo → (served via GitHub Pages — clicking the link in GitHub's UI to sample_report.html would only show source).

The demo shows the stats grid, trend sparkline, failure cards with embedded screenshots + step lists, and the collapsed Passed section.

Integrations

mk-qa-master doesn't bundle third-party SDKs — it stays a pure test-execution + analysis layer. Real QA workflows are composed by running multiple MCP servers side-by-side in the same client config; Claude orchestrates the chain across servers. There's no MCP-to-MCP RPC — each server is independent, the AI client is the conductor.

The pairings below are the ones that complete the loop most often:

Honorable mention — Google Drive MCP: pairs with Google-Sheet-based TC management (read TCs from a sheet → generate_test → write status back).

Composing in your client config

All five run as separate processes alongside mk-qa-master:

{
  "mcpServers": {
    "mk-qa-master": { "command": "python", "args": ["-m", "mk_qa_master.server"], "env": { "QA_RUNNER": "maestro" } },
    "atlassian":       { "command": "npx", "args": ["-y", "@atlassian/mcp"] },
    "slack":           { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-slack"] },
    "github":          { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"] }
  }
}

Then a single prompt walks the chain:

"Run the checkout suite. For each failure, open a JIRA in project QA with the RIDER format and the screenshot attached. Post the HTML report to #qa-bots when done."

Why this matters: mk-qa-master stays focused on the test loop (analyze → generate → run → coach). JIRA / Slack / Sentry are entire domains with their own dedicated servers — bolting them into this one would dilute the scope, duplicate auth handling, and force every user to inherit dependencies they may not want.

本 repo 不打包任何第三方 SDK——維持「測試執行 + 分析」單一職責。實務上 QA 工作流是多個 MCP server 並存、由 Claude 編排跨 server 的 tool chain達成的。範例配套：JIRA / Slack / GitHub / Sentry / Filesystem 各自獨立 MCP server，配上 mk-qa-master 拼出完整測試管線。

Publishing (maintainer-only)

Releases ship to PyPI via Trusted Publishing — no API tokens stored in the repo. The flow:

1. Bump version = "x.y.z" in pyproject.toml (via a normal PR — main is branch-protected).
2. After merge, tag main and push:
   bash

   Copy

   git tag -a vX.Y.Z -m "vX.Y.Z — short summary"
   git push origin vX.Y.Z
   

3. Create a GitHub Release for that tag (gh release create vX.Y.Z ...).
4. The release event fires .github/workflows/publish.yml → builds sdist + wheel → uploads to PyPI.

One-time PyPI setup (must be done once before the first publish works):

• Sign in at https://pypi.org → enable 2FA.
• Project page → Settings → Publishing → add a pending publisher with:
  • Owner: kao273183
  • Repository: mk-qa-master
  • Workflow filename: publish.yml
  • Environment name: pypi

After the first successful run, PyPI auto-promotes the pending publisher to a trusted one and subsequent releases authenticate via OIDC.

The workflow refuses to publish if the release tag doesn't match pyproject.version, which catches "tagged but forgot to bump" mistakes before they hit PyPI.

Support the project ☕

mk-qa-master is built and maintained solo on nights and weekends. If it saved you time or shaped how your team thinks about AI-driven QA, a coffee keeps the late-night Maestro debugging sessions going:

Your support funds: keeping this repo free + actively maintained, more device variants for Maestro testing (real iPhones / Android tablets / BlueStacks), recorded tutorials for the QA community, and the next 2am bug hunt.

No ads, no sponsorships, no enterprise upsell — just the work.

Contributing

This repo is maintained solo. Ideas and bug reports are very welcome — please open an Issue or start a Discussion. I read every one and will implement what fits the project's direction.

External pull requests are auto-closed. Not because contributions aren't appreciated, but because keeping the codebase coherent under a single voice matters more here than the throughput a multi-contributor model would bring. If you really want a specific change, an Issue describing the problem gets you further than a PR.

本 repo 由我一人維護。歡迎透過 Issue / Discussion 提想法或回報問題，我會親自評估並實作。外部 PR 會自動關閉——不是不歡迎貢獻，而是想保持程式碼風格與走向一致。

License

MIT © 2026 Jack Kao — see LICENSE (中文翻譯參考: LICENSE.zh-TW.md; the English version is authoritative).

In plain English: you can use this for anything (personal projects, commercial work, modifications, redistribution). The only ask is that you keep the copyright + license notice in any copy you ship. There's no warranty — use at your own risk.