https://github.com/arisrhiannon/vibecheck
Offline, zero-dependency, no-AI "safe to ship?" gate for vibe-coded apps — finds the security mistakes AI coding agents make (secrets, unprotected routes, eval-RCE, SQLi, CORS, JWT, leaked keys). CLI + MCP, agent-runnable.
https://github.com/arisrhiannon/vibecheck
agent ai-generated-code ci devsecops llms-txt mcp owasp sast secrets security ship-gate typescript vibe-coding zero-dependency
Last synced: 1 day ago
JSON representation
Offline, zero-dependency, no-AI "safe to ship?" gate for vibe-coded apps — finds the security mistakes AI coding agents make (secrets, unprotected routes, eval-RCE, SQLi, CORS, JWT, leaked keys). CLI + MCP, agent-runnable.
- Host: GitHub
- URL: https://github.com/arisrhiannon/vibecheck
- Owner: ArisRhiannon
- License: mit
- Created: 2026-05-31T03:14:47.000Z (1 day ago)
- Default Branch: main
- Last Pushed: 2026-05-31T04:49:26.000Z (1 day ago)
- Last Synced: 2026-05-31T05:10:57.089Z (1 day ago)
- Topics: agent, ai-generated-code, ci, devsecops, llms-txt, mcp, owasp, sast, secrets, security, ship-gate, typescript, vibe-coding, zero-dependency
- Language: TypeScript
- Size: 95.7 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
- Agents: AGENTS.md
Awesome Lists containing this project
README
# vibecheck
[](https://github.com/ArisRhiannon/vibecheck/actions/workflows/ci.yml)
[](LICENSE)
A fast, **agent-native** "safe to ship?" gate for vibe-coded apps. It parses your **JS/TS/JSX/TSX**
(`@babel/parser`), **Python** (the stdlib `ast`), and **Go** (`go/parser`) with **real parsers** and uses **taint analysis**
(inter-procedural for JS/TS — return-taint + param→sink summaries, within a file and across files by
imported name) to flag the security classes AI coding agents get wrong — committed secrets, SQL
injection through *abstracted* raw-query APIs, XSS, SSRF, path traversal, command injection, insecure
deserialization, weak JWT/CORS/cookies — and ranks every finding by **confidence** so an agent can fix
the real ones and ignore the noise.
```sh
vibecheck . # human report (severity + confidence)
vibecheck . --ci # exit 1 only on high-confidence (taint-backed) issues
vibecheck . --json # machine-readable findings for agents / CI
```
## What it is — and what it is not (read this)
vibecheck is **not** a replacement for [Semgrep](https://semgrep.dev) or
[CodeQL](https://codeql.github.com). Those are deeper, broader, multi-language engines and you should
run them for full coverage. vibecheck aims to be **better on one narrow, measurable axis**: a
**low-false-positive, taint-backed gate for the AI-vibe-coding failure classes that runs inside agent
loops and pre-commit in milliseconds**, with **published precision/recall** so you can trust the
`--ci`/MCP signal. Use it *alongside* the big engines, not instead of them.
| | vibecheck | Semgrep | CodeQL |
|--|--|--|--|
| Parsing | real AST (Babel JS/TS/JSX + Python `ast`) | real, many langs | real, many langs |
| Data-flow | inter-procedural (return-taint + param→sink; intra-file + cross-file by import) | taint (Pro) | full inter-procedural |
| Languages | **JS/TS/JSX/TSX + Python + Go** | many | many |
| Speed / infra | ms, local, no account | fast | slower, CI-oriented |
| Agent-native (MCP, confidence gating) | **yes, first-class** | partial | no |
| Breadth of rules | small, focused | 2000+ | huge |
If you only adopt one general SAST, adopt Semgrep or CodeQL. Adopt vibecheck as the **fast agent/CI
pre-flight** that won't drown an agent in false positives.
## Measured quality (not claimed)
Against a labeled benchmark of **77 cases** across JS/TS, Python **and** Go (vulnerable + safe + deliberately
tricky-safe), the core detectors score (see [`METRICS.md`](METRICS.md), reproduce with `bun benchmark/run.ts`):
- **Precision 100%**, **Recall 100%**, **F1 100%** on the corpus.
The tricky-safe cases that produce **zero false positives** include: parameterized queries, tagged-
template SQL, numeric-coerced and schema-validated input, ORM/RegExp `.exec()`, Supabase **anon** /
Stripe **publishable** keys, hardened cookies, allow-listed CORS, and pinned JWT algorithms — exactly
the patterns a regex linter trips on. This benchmark is curated; for a **real-world** measurement (pinned
OSS repos, manually triaged), see [`docs/CORPUS.md`](docs/CORPUS.md) — which also drove a precision fix
(fixed-prefix relative redirects are no longer flagged).
## Confidence (the anti-false-positive-loop design)
Every finding has a **confidence**:
- `high` — a user-input **source provably flows into the sink** (taint-backed), or a deterministic fact
(committed secret, JWT `none`). These fail `--ci` and are what the MCP `scan` tool returns by default.
- `medium` — a dangerous sink on a non-literal value with no proven source (e.g. `eval(x)`).
- `review` — a structural smell that needs a human (e.g. a route with no visible auth). **Excluded from
`--ci` and from the agent loop by default**, so agents never chase phantom work. Add `--all` to include them.
## Install & use
```sh
bun add -d vibecheck # or: npm i -D vibecheck (Node >= 20)
vibecheck . --ci
vibecheck explain VC-SQLI
vibecheck mcp # MCP stdio server exposing a `scan` tool (high-confidence by default)
```
Agents: see [`AGENTS.md`](AGENTS.md) — run `vibecheck . --ci` before declaring a task done and fix every
high-confidence finding.
> JS/TS scanning needs nothing extra. **Python** scanning requires **`python3` on PATH**; **Go** scanning
> requires a **`go` toolchain on PATH** (the analyzer is compiled once and cached). If a runtime is absent,
> those files are silently skipped.
## Rules (exactly what is implemented + benchmarked)
Taint-backed: `VC-RCE-EVAL`, `VC-RCE-CHILD-PROCESS`, `VC-SQLI`, `VC-XSS-REACT`, `VC-XSS-DOM`, `VC-SSRF`,
`VC-PATH-TRAVERSAL`, `VC-OPEN-REDIRECT`. AST config: `VC-CORS-WILDCARD`, `VC-JWT-NONE`,
`VC-JWT-UNPINNED`, `VC-COOKIE-INSECURE`, `VC-STACK-EXPOSURE`. Provenance/secrets: `VC-SECRET-*` (8),
`VC-ENV-COMMITTED/DRIFT/MISSING`, `VC-NEXT-PUBLIC-SECRET`, `VC-SUPABASE-SERVICE-ROLE`. Advisory:
`VC-ROUTE-NO-AUTH` (review), `VC-INPUT-NO-VALIDATION`. **Python** (`VC-PY-*`): `VC-PY-RCE`,
`VC-PY-CMDI`, `VC-PY-SQLI`, `VC-PY-DESERIALIZE`, `VC-PY-YAML`, `VC-PY-SSTI`, `VC-PY-OPEN-REDIRECT`,
`VC-PY-PATH`. **Go** (`VC-GO-*`): `VC-GO-CMDI`, `VC-GO-SQLI`, `VC-GO-PATH`, `VC-GO-OPEN-REDIRECT`,
`VC-GO-SSRF`. `vibecheck explain ` prints the fix for each.
## Limitations (honest)
- **JS/TS/JSX/TSX + Python + Go** (Python needs `python3`, Go needs a `go` toolchain on PATH). More
languages are roadmap (each via its own real parser, never hand-rolled).
- **Taint scope:** JS/TS taint is **inter-procedural** — function summaries carry **return-taint** and
**parameter→sink** reachability, resolved **within a file and across files by imported name** (1-hop,
sanitizers respected). Not tracked (false negatives): aliased/namespace/re-exported imports, helper
chains beyond a few hops, methods, destructured params, and names defined in **multiple** files
(treated as ambiguous and skipped — to avoid false positives). Python and Go are intra-procedural
(Go does not track multi-return assignments like `x, _ := f(src)`).
- Config/secret rules are pattern-based where AST adds no value.
- A high-signal gate and early-warning — **not a proof of security**. Pair it with Semgrep/CodeQL and review.
## Config — `.vibecheck.json`
```json
{ "ignoreRules": ["VC-INPUT-NO-VALIDATION"], "allowPaths": ["test/**"], "failSeverity": "high" }
```
## License
MIT © 2026 Aris Rhiannon — see [LICENSE](LICENSE).