test-genie-mcp
Built for vibe coders: one command, get a prioritized list of what's actually broken about your project.
Self-healing test automation for iOS, Android, Flutter, React Native and Web apps — as an MCP server.
에이전트 강화
게시자의 자체 명령입니다. PlugMCP 엔드포인트가 아닙니다.
Host a copy with our keys and logs.
npx -y [email protected]Claude Code에는 로컬 명령용 설치 링크가 없습니다. 아래 명령을 복사해 터미널에서 실행합니다.
claude mcp add io.github.muse-code-space-test-genie -- npx -y test-genie-mcp@2.0.3터미널에서 실행합니다.
{ "mcpServers": { "io.github.muse-code-space-test-genie": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}claude_desktop_config.json
원격 URL인가요, 내 컴퓨터에서 실행하는 명령인가요? 두 설치 방식의 차이
PlugMCP에서 실행
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
실행게시자 제공
v3.1.1 — vibe-check + honest auto-fix. One MCP call, ~30 seconds: race conditions + security issues + memory leaks + logic errors + perf smells, prioritized. Stays on your machine, no telemetry. Pass
autoFix: truefor the small, safe mechanical fixes (weak-hash, simpleMath.randomassignment) — backup + syntax-validate + rollback-on-syntax-fail. For test-verified application of harder fixes, use v3.0.0's iterate-fix loop.
You don't read the docs. You open the project, talk to Claude, and want a verdict. Here it is:
In Claude (with test-genie-mcp installed — setup):
/vibe-check /Users/me/my-app
Claude calls diagnose_project under the hood. ~30 seconds later you see:
# vibe-check report
- Project: /Users/me/my-app
- Platform: web
- Findings: 11 total — 4 critical, 4 high, 1 medium, 1 low
- Estimated fix time: ~85 min
## Top 5 issues
### 1. [CRIT] Hardcoded AWS access key id found in source
- File: `server.js:7`
- Category: security / secret (CWE-798)
- Confidence: 95%
- Fix: Move the value to an env var, gitignore the config, rotate the leaked key.
### 2. [CRIT] SQL string built by concatenating user input
- File: `server.js:21`
- Category: security / injection (CWE-89)
- Fix: Use parameterized queries (`db.query("... WHERE id = ?", [id])`).
### 3. [HIGH] useState setter called after await without mount guard
- File: `UserProfile.tsx:16`
- Category: race-condition / react-setstate-after-await (CWE-362)
- Confidence: 78%
- Fix: Use AbortController and check signal.aborted before calling setters.
… (top 5 shown — full list at output: "detailed")
## Next steps
1. Address the critical / high findings above.
2. Re-run diagnose_project after fixing to confirm convergence.
3. Use run_iterative_fix_loop for test-driven verification of each fix.
If any finding is autoFixable: true and is at high/critical severity, the diagnose_project call accepts autoFix: true to apply the mechanical replacement directly (with backup + syntax validation — see SAFETY.md for the exact guards). The v3.1.1 honest scope is narrow: weak hash (createHash('md5'|'sha1') → createHash('sha256')) and standalone Math.random() in security-sensitive files. For broader/structural fixes (race conditions, eval, exec injection) run run_iterative_fix_loop separately — it re-runs tests and auto-rolls-back on regression.
The bottleneck in mobile + cross-platform test automation isn't writing tests — it's the loop between a failing test and a passing test. test-genie closes that loop:
failing test → analyzer flags issue → fix proposed → dry-run + syntax check →
applied with backup → affected tests re-run → regression check → loop or stop
This full loop is the run_iterative_fix_loop tool. The diagnose_project autoFix: true path in v3.1.1 covers a strict subset — backup + dry-run + syntax-validate + apply, without re-running tests (so no test-regression rollback in that path). Use the right tool for the job — and see SAFETY.md for the exact guards on each.
Other tools (Detox, Maestro, Playwright, xcodebuild test) run tests. test-genie runs tests and drives the fix until the bar is met or it can no longer make progress — without you scrubbing through stack traces.
# 1. Install
npm install -g test-genie-mcp
# 2. Add to Claude Desktop config (~/.config/claude/claude_desktop_config.json)
{
"mcpServers": {
"test-genie": {
"command": "npx",
"args": ["test-genie-mcp"],
"env": {
"TEST_GENIE_ALLOWED_ROOT": "/path/to/your/project"
}
}
}
}
# 3. Restart Claude Desktop. From a chat:
# "Run the iterate-fix loop on /Users/me/my-rn-app with autoApply=false"
Expected output (truncated):
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Iterative fix loop f8b3… — PAUSED-FOR-CONFIRMATION
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Iterations completed: 1
Fixes applied: 0
Regressions rolled back: 0
Final tests: 7/10 passing (3 failing)
Pending confirmations (3):
- 71fbe…: Fix: useEffect missing cleanup for setInterval (confidence: 85)
- 92ad1…: Fix: Force-unwrap on possibly-undefined name (confidence: 85)
- …
Resume token: f8b3…
Re-call with autoApply: true (or resumeToken: "f8b3…") to actually patch the files.
The flows below describe the
run_iterative_fix_looppath (v3.0 headline) — full detect → propose → dry-run → apply-with-backup → re-run-tests → rollback-on-regression. Thediagnose_project autoFixpath in v3.1.1 is the narrower mechanical-replacement-only path; see SAFETY.md §4 for what that one actually touches.
A team adds setInterval(...) in a useEffect and forgets cleanup. test-genie's detect_memory_leaks flags it, suggest_fixes proposes return () => clearInterval(id) (src/tools/fixing/suggestFixes.ts:169-179), the loop dry-runs the patch through the TS compiler, applies with backup, re-runs only the affected snapshot test, confirms 100% pass, stops. Before: 1 failing snapshot. After: 0 failing, 1 fix applied, 1 backup at .test-genie-backups/.
dispose() automationAnimationController left undisposed. test-genie sees the missing dispose() override, generates a Dart @override dispose() { controller.dispose(); super.dispose(); } block (suggestFixes.ts:214-217), runs dart analyze on the patched file, applies, re-runs flutter test, converges.
self.timer = Timer.scheduledTimer(...) { _ in self.tick() } — rule-based detector flags closure self-capture, fixer rewrites to [weak self] _ in guard let self = self else { return }; self.tick() (suggestFixes.ts:239-242). If swiftc is on PATH the syntax check is real; otherwise test-genie reports "downgraded validation" so you know.
┌────────────────────┐
│ collect tests │ (run_scenario_test / supplied list)
└─────────┬──────────┘
│
pass-rate ≥ threshold? ── yes ──▶ SUCCESS
│ no
▼
┌────────────────────┐
│ detect issues │ memory + logic analyzers
└─────────┬──────────┘
│
┌────────────────────┐
│ suggest fixes │ rule-based (default) → LLM (hybrid, optional)
└─────────┬──────────┘
│
┌────────────────────┐
│ dry-run + syntax │ TS compiler API / platform compiler / brace check
└─────────┬──────────┘
│
┌────────────────────┐
│ apply with backup │ per-file `.test-genie-backups/`
└─────────┬──────────┘
│
┌────────────────────┐
│ re-run tests │ regression? yes → auto-rollback
└─────────┬──────────┘
│
▼
loop (≤ maxIterations, ≤ totalTimeout)
See docs/ITERATE_FIX_LOOP.md for a sequence diagram and the full safety-guard list.
mode legend in docs/SIMULATION_VS_REAL.md.
Plus 4 resources (test-genie://iteration-logs, …/test-history/{path}, …/iteration-logs/{loopId}, …/applied-fixes/{path}) and 3 prompts (full-test-pipeline, diagnose-failure, vibe-check).
Race conditions (detect_race_conditions / diagnose_project):
v3.1.1 honesty audit:
useEffect-no-abortandforEach-awaitwere previously advertised as auto-fixable. They are not — wrapping withAbortControlleror rewriting toPromise.all(arr.map(...))changes behavior we can't verify statically. They are now report-only. See SAFETY.md.
Security (detect_security_issues / diagnose_project):
v3.1.1 honesty audit:
.env/Math.random(general)/yaml.loadwere previously advertised as auto-fixable. They were either too risky to rewrite blindly or no strategy shipped — flipped to report-only. See SAFETY.md §5.
This is a "catch the obvious stuff in 30s" filter, not Snyk / Semgrep / a full SAST tool. We don't catch:
db.query, the regex won't connect the dots. A real SAST traces taint across the call graph. Roadmap: ts-morph reference walking for top-N entry points.npm audit's job, and bundling a stale advisory list would lie. Run npm audit --json in parallel if you want dep-CVE coverage.Math.random() named getNonce won't fool us; a properly-named crypto.randomBytes used with a tiny entropy budget will.securityAnalyzer.SECRET_PATTERNS. PR welcome.run_stress_test / run_simulation, not static analysis.If you want deeper coverage on top of vibe-check: feed the findings into run_iterative_fix_loop for test-verified application, or escalate to Snyk / Semgrep / GitHub Advanced Security for compliance use cases.
If your goal is "before I commit, what's broken?", vibe-check wins on latency. If your goal is "compliance + supply chain audit", use the dedicated tools.
autoApply: false (the default) and use it as a fix-proposal generator only.Scope note:
diagnose_project autoFix: truerolls back on syntax-validate failure (applyFix.ts:185-202) but does not re-run tests, so it cannot detect test regressions. For test-driven rollback userun_iterative_fix_loop. See SAFETY.md §2.4.
test-genie uses tools like Jest, Detox, and xcodebuild test under the hood — it sits at the orchestration layer, not the test-runner layer.
-typecheck mode. If the compiler isn't on PATH, we fall back to brace-balance validation and surface downgraded: true in the result. Install swiftc / kotlinc / javac / dart for real validation.strategy: 'hybrid' only kicks LLM in when rule-based confidence is below threshold. Without an API key the loop is rule-based-only — no failure.$TEST_GENIE_STORAGE_DIR (defaults to ~/.test-genie-mcp). Not synced across machines.run_simulation returns plausible anomalies, not real ones. Use run_scenario_test (hybrid) for real-device runs.run_full_automation still works. The confirmMode / autoFix options are kept for compatibility but autoApply: boolean is the new way — autoApply: true is equivalent to confirmMode: 'auto'.Issues, PRs, and ideas welcome — see CONTRIBUTING.md (TODO). Code lives under src/, tests under tests/. Run npm test before sending a PR.
@MUSE-CODE-SPACE — Yoonkyoung Gong.
MIT — see LICENSE.
AI-powered app test automation for iOS, Android, Flutter, React Native, Web
{ "mcpServers": { "io.github.muse-code-space-test-genie": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}도구
아직 도구 목록이 없습니다. 원격 서버는 첫 상태 확인 후에 도구 목록이 생깁니다.
FAQ
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
Test Genie은(는) stdio를 통해 로컬에서 실행됩니다.
마지막 확인: 2026-11-09. AI-powered app test automation for iOS, Android, Flutter, React Native, Web