CodeInspectus, by Synvoya

A local-first, privacy-preserving security MCP server. Any AI coding agent (Claude Code, Cursor, Codex, Windsurf, Cline, Aider) can invoke CodeInspectus to scan AI-generated / "vibe-coded" code for real vulnerabilities, map findings to compliance frameworks as honest code-level coverage, and drive a scan → fix → rescan loop — fully on your machine, with no account and zero network egress at scan time.

Reproduce the proof: the shipped intentionally vulnerable fixture produces 18 normalized findings across all four engines with v0.3.1 (4 critical, 8 high, 5 medium, 1 low). Inspect the fixture, read the full scanner-derived report, or run the 17-case eval suite. Dependency findings can change as the local Trivy database updates; the report records the exact engine and database versions used.

If CodeInspectus is useful, star the repository so other AI-app builders can find it.

CodeInspectus orchestrates three best-in-class OSS engines behind one normalized, CWE-keyed schema, and adds its own AI-code-specific checks that generic scanners miss:

• Opengrep — SAST / OWASP Top 10 (SARIF)
• Gitleaks — secrets
• Trivy — dependency CVEs (SCA), IaC misconfig, secrets, license, SBOM
• CodeInspectus AI checks — client-side secret/bundle exposure, Supabase RLS / inverted-auth (the CVE-2025-48757 class), prompt-injection sinks, client-writable user_metadata authorization, and unsanitized model/user output rendered via dangerouslySetInnerHTML (XSS / LLM05)

CodeInspectus bundles the official, SHA-pinned engine binaries and calls them as local subprocesses. It does not fork them.

Why CodeInspectus?

AI-generated apps often ship with security mistakes that generic scanners miss: exposed client-side secrets, weak Supabase auth patterns, unsafe HTML rendering, prompt-injection sinks, and insecure agent/tool integrations.

CodeInspectus combines proven local scanners with AI-app-specific rules, then exposes the workflow through an MCP server so coding agents can scan, explain, and help fix issues before shipping.

Install

Prerequisites: Node.js ≥18, and cosign on your PATH for install-engines. Signature verification is fail-closed: Opengrep and Trivy will not pin without cosign — install-engines exits non-zero for them — so install it first (brew install cosign). Gitleaks verifies by checksum and needs no cosign.

# Register once per machine with your agent (see "Client registration"), then:
npx codeinspectus install-engines

install-engines is the only step that touches the network. It downloads the engine binaries from their verified GitHub release URLs, checks the publisher signature/checksum, computes each binary's SHA256, and records it in engines.lock.json. It also fetches the offline Trivy vulnerability-DB snapshot into ~/.codeinspectus/. After this, scans perform zero network I/O.

If a Trivy DB was installed before 0.3.2, scan output tells your agent that CVE rescan tracking is not yet enabled. The agent should run npx codeinspectus install-engines once; this re-fetches the DB through the verified install path and records its provenance. Until then, vanished CVEs conservatively report not_rechecked; current scan findings remain complete and unaffected.

Re-verify your pinned binaries any time:

npx codeinspectus verify-engines

An MCP server is installed once per machine and shared across all your projects — it is not a per-repo npm install dependency.

Client registration

The server command is the same across clients, but each client uses its own configuration format. Clients with JSON MCP configuration use:

{
  "mcpServers": {
    "codeinspectus": { "command": "npx", "args": ["-y", "codeinspectus"] }
  }
}

For Codex, choose one registration method:

# Codex CLI
codex mcp add codeinspectus -- npx -y codeinspectus

• Codex app or IDE extension: open Settings → MCP servers → Add server, choose STDIO, set the command to npx and arguments to -y, codeinspectus, then restart.
• Codex configuration file: add this to global ~/.codex/config.toml or a trusted project's .codex/config.toml:

[mcp_servers.codeinspectus]
command = "npx"
args = ["-y", "codeinspectus"]
tool_timeout_sec = 600

Codex defaults MCP tool calls to 60 seconds. CodeInspectus runs multiple security engines concurrently and permits up to five minutes per engine, so 600 seconds avoids premature client timeouts on larger repositories. This is a Codex client timeout only; it does not change engine limits or other clients' configurations.

Optional: drop in the ready-made agent-rules/ so your agent auto-runs the scan → fix → rescan loop.

MCP clients that support server instructions, including Codex, also receive the safe workflow automatically: show findings first, obtain granular approval before fixes, and rescan before claiming an issue is resolved. The agent-rule files remain useful when you want the same policy persisted explicitly in a repository.

Tools

CodeInspectus never edits or deletes your source code or repository — it reads and reports; your agent applies the fixes. It stores engine data and scan history under ~/.codeinspectus; the only file it writes is an optional SBOM — to a managed directory by default, or a path you choose (see codeinspectus_generate_sbom).

Each scan also reports a read-only git-safety state: if there's no git repo or uncommitted changes, it recommends creating a checkpoint before fixes — your agent runs git only with your approval; the tool never does.

Honest claims (please read)

• "No egress" is precise: zero egress at scan time. Engine binaries and the initial Trivy DB are fetched at install time from verified sources, with SHA256 verification. The scanner functions with the network unplugged. There is no telemetry, ever.
• Supply-chain pinning is mandatory. Trivy was supply-chain-compromised twice in early 2026; every engine binary is SHA-pinned in engines.lock.json and its hash is verified before execution. CodeInspectus refuses to run an unpinned or mismatched binary.
• Secret values are redacted in all output — type + location + a redacted preview only.
• Compliance = code-level control coverage, never certification. CodeInspectus reports "X of N code-visible controls have findings", with the code-visible subset as the explicit denominator, plus a standing disclaimer. It never emits "you are X% compliant" or "you pass [framework]". The severity-weighted posture score is a separate view and is not a percent-compliant figure. Essential Eight especially: only ~1 of 8 mitigations (Patch Applications) is code-evidenced — this is not an Essential Eight assessment.
• Prompt-injection detection is heuristic and immature — those findings are worded "potential …" and marked medium confidence.
• Client-side authorization that trusts user_metadata is flagged (ci-ai-client-metadata-authz). CodeInspectus detects an authorization decision that reads client-writable Supabase user_metadata — e.g. if (user.user_metadata.role === 'admin') — at high severity, medium confidence (CWE-639). user_metadata is editable by the signed-in user themselves (Supabase's /auth/v1/user endpoint), so anyone can self-assign role: 'admin'; gate privileged logic on the server-controlled app_metadata.role instead. Detection is intrafile (inline + split-variable/destructured); it does not yet trace cross-file or whole-object-alias flows (planned) — see the good-first-issue. It also catches the related footgun: a Supabase service_role key value in client-reachable code (critical), and a service_role key behind a client-exposed env prefix such as NEXT_PUBLIC_… (high).
• Unsanitized model or user output rendered as raw HTML is flagged (ci-ai-llm-output-dangerous-html). CodeInspectus detects untrusted request input or LLM/model output flowing into dangerouslySetInnerHTML without sanitization — a direct XSS sink (CWE-79/116; OWASP LLM05 on the model-output path), high severity, medium confidence; wrapping the value in DOMPurify.sanitize(...) silences it. It does not yet trace untrusted values arriving via component props, database rows, or template data (planned).

Language support

Plainly, what runs on what. The commodity engines are broad; the CodeInspectus AI-code checks (the moat) are JavaScript/TypeScript-focused today — more languages are planned. So on a Python/Go/Rust/etc. repo you still get full secrets, dependency, IaC and SBOM coverage (and Python SAST), but the AI-code-specific checks won't fire. This is stated so you don't infer coverage that isn't there.

Compliance frameworks (code-visible subset)

NIST CSF 2.0 · ISO/IEC 27001:2022 · SOC 2 · CIS Controls v8.1 · Essential Eight (Patch Applications only) · OWASP Top 10 (2021) · OWASP LLM Top 10 (2025). MITRE ATT&CK techniques are shown as related-adversary context only, never as a coverage score.

Compliance mappings are AI-drafted, reviewed by a cybersecurity practitioner (Synvoya) — code-level coverage only, not an audit or certification. Community review welcome. The CWE→control mappings are self-audited with per-mapping confidence and an open community-verification process — see docs/COMPLIANCE-RATIONALE.md and CONTRIBUTING.md. Essential Eight is not a coverage view: only Patch Applications is code-evidenced (~1 of 8) — this is not an Essential Eight assessment.

How it works

agent → codeinspectus_scan → [Opengrep | Gitleaks | Trivy] + AI checks
      → SARIF normalize → dedup (incl. Trivy⨯Gitleaks secret overlap)
      → CWE-keyed findings → compliance map → compact JSON + summary
ALL LOCAL. NO NETWORK EGRESS AT SCAN TIME.

Example reports

• Reproducible v0.3.1 scan of the shipped vulnerable fixture

Trademark

"CodeInspectus" is the name of this free, open-source project (npm codeinspectus, codeinspectus.com). "Code Inspect" is a descriptive phrase in a crowded namespace; registry availability is not trademark clearance, and the name is not claimed as a trademark.

Development

npm install
npm run build      # tsc --noEmit && tsup  (must compile clean)
npm run eval       # ≥10 evals against fixtures/vulnerable-app
npm run inspector  # npx @modelcontextprotocol/inspector node dist/index.js

How this repository is generated (an auditable, allow-list seed) and built end-to-end: docs/BUILD.md.

Contributing

CodeInspectus is a solo, free, open-source project, built and maintained by one cybersecurity practitioner under the Synvoya name. There is no company behind it and nothing to sell — which is exactly why outside eyes matter. Independent review is genuinely wanted, not a courtesy line. If you work in security, your scrutiny is the contribution.

Two areas where review helps most:

• Compliance CWE→control mappings. These are AI-drafted, then policy-reviewed by the maintainer — they are NOT independently verified. Every mapping is tracked through three explicit states: AI-drafted → maintainer-policy-reviewed → community-verified. Today almost everything sits in the first two; the community-verified count is 0 of 96, and that is reported honestly rather than hidden. Moving a mapping to community-verified takes evidence (a quote from the control's primary source + your basis) — the bar and process are in CONTRIBUTING.md; the per-mapping rationale and confidence live in docs/COMPLIANCE-RATIONALE.md.
• Detection rules (detection-db/**, src/ai-checks/**). New rules, precision fixes, and false-positive reports are all welcome. The merge bar is precision: a fixture proving the true positive, and a near-miss fixture proving the rule does not over-fire. Details in CONTRIBUTING.md.

What CodeInspectus claims — and what it deliberately does not — is written down so you can check it before trusting a number: the standing compliance disclaimer (in the Compliance frameworks section above and in docs/COMPLIANCE-RATIONALE.md) and the three-state honesty metric. If something reads as over-claiming, that is a bug — please open an issue.

Workflow: fork → branch → PR; the maintainer reviews and merges (external contributors don't push directly). — Synvoya (the maintainer, a cybersecurity practitioner)

Good first contributions

• Trace tainted component props into dangerouslySetInnerHTML
• Detect model output passed to dynamic execution or shell sinks
• Detect unguarded Next.js admin API routes
• Community-verify one CWE to OWASP Top 10 mapping
• Community-verify one CWE to SOC 2 or ISO 27001

Browse all good first issues.

Changelog

Per-version release notes live in CHANGELOG.md.

Licenses

CodeInspectus: MIT. Bundled engines: Opengrep (LGPL-2.1), Gitleaks (MIT, CLI only), Trivy (Apache-2.0) — all permissive for bundling the compiled binaries.