test-genie-mcp
Self-healing test automation MCP server that diagnoses and fixes issues in mobile and web apps, including a one-command vibe-check to prioritize broken functionality.
README
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.
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.
Vibe coders quickstart
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.
Why test-genie?
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.
5-minute Quickstart
# 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.
Real use cases
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.
1. React Native memory-leak self-healing
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/.
2. Flutter widget dispose() automation
AnimationController 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.
3. iOS retain-cycle (closure capture)
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.
How the iterate-fix loop works
┌────────────────────┐
│ 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.
Tools (23)
| # | Tool | Mode |
|---|---|---|
| 1 | analyze_app_structure |
real |
| 2 | generate_scenarios |
real |
| 3 | create_test_plan |
real |
| 4 | run_scenario_test |
hybrid |
| 5 | run_simulation |
simulated |
| 6 | run_stress_test |
hybrid |
| 7 | detect_memory_leaks |
real |
| 8 | detect_logic_errors |
real |
| 9 | suggest_fixes |
real |
| 10 | confirm_fix |
real |
| 11 | apply_fix |
real |
| 12 | rollback_fix |
real |
| 13 | run_full_automation |
hybrid |
| 14 | run_iterative_fix_loop (v3.0 headline) |
hybrid |
| 15 | generate_report |
real |
| 16 | get_pending_fixes |
real |
| 17 | get_test_history |
real |
| 18 | analyze_performance |
real |
| 19 | analyze_code_deep |
real |
| 20 | generate_cicd_config |
real |
| 21 | diagnose_project (v3.1 headline — vibe-check) |
real |
| 22 | detect_race_conditions |
real |
| 23 | detect_security_issues |
real |
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).
What vibe-check catches
Race conditions (detect_race_conditions / diagnose_project):
| Pattern | Language | Severity | Auto-fixable (v3.1.1) |
|---|---|---|---|
useState setter called after await without mount guard |
TS/JS/React | high | no (structural) |
useEffect with async fetch, no AbortController/cleanup |
TS/JS/React | high | no (structural) |
arr.forEach(async ...) (silent fire-and-forget) |
TS/JS | medium | no (ordering-sensitive) |
Adjacent fetches without Promise.all / sequencing |
TS/JS | medium | no |
TOCTOU: existsSync then readFileSync without lock |
TS/JS Node | medium | no |
| Non-atomic counter increment in async context | TS/JS | low | no |
@Published mutation outside @MainActor |
Swift | medium | no |
Concurrent DispatchQueue writes without .barrier |
Swift | medium | no |
MutableStateFlow mutated off Dispatchers.Main |
Kotlin | medium | no |
Flow collected without flowOn |
Kotlin | low | no |
Goroutine + shared map without sync.Mutex |
Go | high | no |
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):
| Pattern | Severity | CWE | Auto-fixable (v3.1.1) |
|---|---|---|---|
| Hardcoded AWS / Stripe / GitHub / Google / Slack token | critical / high | CWE-798 | no (rotate) |
| Hardcoded JWT secret literal | high | CWE-798 | no |
| API token in URL query string | high | CWE-200 | no |
.env file present but not gitignored |
high | CWE-538 | no (rotation must follow) |
SQL string concat with req.params / req.body |
critical | CWE-89 | no |
innerHTML / dangerouslySetInnerHTML with dynamic value |
high | CWE-79 | no |
eval() / new Function() with non-literal |
critical | CWE-95 | no |
Math.random() in security-sensitive file, standalone assignment |
high | CWE-338 | yes (crypto.randomInt) |
Math.random() mixed into arithmetic |
high | CWE-338 | no (semantic) |
createHash('md5'|'sha1') in security-keyword file |
high | CWE-327 | yes ('sha256') |
createHash('md5'|'sha1') elsewhere |
medium | CWE-327 | no (below severity floor) |
child_process.exec with user-input template literal |
critical | CWE-78 | no |
fetch(req.query.url) (SSRF) |
high | CWE-918 | no |
CORS * origin + Allow-Credentials: true |
high | CWE-942 | no |
Cookie set without httpOnly / secure / sameSite |
low | CWE-1004 | no |
yaml.load without safe schema |
medium | CWE-502 | no |
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.
What vibe-check misses (honest list)
This is a "catch the obvious stuff in 30s" filter, not Snyk / Semgrep / a full SAST tool. We don't catch:
- Cross-file data-flow. If user input flows through three files before reaching a
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. - Vulnerable transitive deps. We don't query npm advisories — that's
npm audit's job, and bundling a stale advisory list would lie. Runnpm audit --jsonin parallel if you want dep-CVE coverage. - Race conditions across processes. We catch in-process JS / Swift / Kotlin / Go races. Distributed races (lock ordering across services, DB transactions) need different tooling.
- Type-correct but logic-broken code. The analyzer is syntactic, not semantic. A
Math.random()namedgetNoncewon't fool us; a properly-namedcrypto.randomBytesused with a tiny entropy budget will. - Custom secret formats. Internal company tokens with unique prefixes need a regex you can add to
securityAnalyzer.SECRET_PATTERNS. PR welcome. - Real-time / dynamic issues. Memory leaks under load, network timeouts, slow renders mid-interaction — those need
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.
vibe-check vs alternatives
| vibe-check (test-genie) | Snyk | Semgrep | GitHub Advanced Security | |
|---|---|---|---|---|
| Runs locally | yes | hybrid (cloud) | yes | no (cloud) |
| Telemetry-free | yes (zero network calls) | no | partial | no |
| Fix loop integration | yes (run_iterative_fix_loop) |
no | no | no |
| Race-condition detection | yes (JS/Swift/Kotlin/Go) | no | partial | partial |
| Cross-file taint flow | no (roadmap) | yes | yes | yes |
| Setup time | none (already installed if test-genie is installed) | account + auth | install + ruleset | repo-level enable |
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.
When NOT to use test-genie
- Production-gate test runs. test-genie is built for the development feedback loop. For shipping decisions, use a proper CI that you control end-to-end.
- Code your team must hand-review every line of. The loop's job is to propose and apply fixes; if every fix needs a human eye, leave
autoApply: false(the default) and use it as a fix-proposal generator only. - No backup / no version control situations. test-genie's auto-rollback is best-effort and requires the per-file backup to exist. Always run inside a git working tree.
Comparison
| test-genie | Detox | Maestro | xcodebuild test | |
|---|---|---|---|---|
| Runs E2E / unit tests | ✅ (via Jest/Detox/etc.) | ✅ | ✅ | ✅ |
| Detects code issues | ✅ rule + LLM | ❌ | ❌ | ❌ |
| Iterative fix loop | ✅ (run_iterative_fix_loop) |
❌ | ❌ | ❌ |
| Auto-rollback on test regression | ✅ inside run_iterative_fix_loop only |
❌ | ❌ | ❌ |
| Auto-rollback on syntax failure | ✅ all apply paths | ❌ | ❌ | ❌ |
| MCP-native (talks to Claude / agents) | ✅ | ❌ | ❌ | ❌ |
| Multi-platform | iOS+Android+Web+Flutter+RN | iOS+Android | iOS+Android | iOS 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.
Known limitations
- Platform syntax check downgrade. For Swift/Kotlin/Java/Dart we try the platform compiler in
-typecheckmode. If the compiler isn't on PATH, we fall back to brace-balance validation and surfacedowngraded: truein the result. Installswiftc/kotlinc/javac/dartfor real validation. - LLM is optional and gated.
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. - Storage is per-machine. Test history / iteration logs live under
$TEST_GENIE_STORAGE_DIR(defaults to~/.test-genie-mcp). Not synced across machines. - Simulated mode is "simulation," not magic.
run_simulationreturns plausible anomalies, not real ones. Userun_scenario_test(hybrid) for real-device runs.
Configuration
| Env var | Default | Purpose |
|---|---|---|
TEST_GENIE_ALLOWED_ROOT |
cwd |
Capability-based path safety — server refuses to read/write outside this root. |
TEST_GENIE_STORAGE_DIR |
~/.test-genie-mcp |
Where scenarios / results / iteration logs live. |
TEST_GENIE_LLM_PROVIDER |
auto-detect | anthropic / openai / none. |
ANTHROPIC_API_KEY |
— | Used when provider = anthropic. |
OPENAI_API_KEY |
— | Used when provider = openai. |
TEST_GENIE_ANTHROPIC_MODEL |
claude-haiku-4-5 |
Override Anthropic model. |
TEST_GENIE_OPENAI_MODEL |
gpt-4o-mini |
Override OpenAI model. |
Migrating from v2.x
run_full_automationstill works. TheconfirmMode/autoFixoptions are kept for compatibility butautoApply: booleanis the new way —autoApply: trueis equivalent toconfirmMode: 'auto'.- Subprocess hardening means platform tools now reject scheme / device / package-name arguments that contain shell metacharacters. If your CI was passing weird-looking values, sanitize them first.
- See CHANGELOG.md for the full breaking-change list + migration recipes.
Roadmap
- LLM-based fix-proposal voting (multiple proposals → pick the best by syntax + retest delta)
- Multi-repo sync (run the loop across N repos in parallel from one MCP call)
- A "watch mode" that runs the loop on file save
- Better Detox / Maestro artifact ingestion (link videos into iteration logs)
Contributing
Issues, PRs, and ideas welcome — see CONTRIBUTING.md (TODO). Code lives under src/, tests under tests/. Run npm test before sending a PR.
Maintainer
@MUSE-CODE-SPACE — Yoonkyoung Gong.
License
MIT — see LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.