CodeInspectus, by Synvoya

A local-first, privacy-preserving security MCP server and CLI. 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.

CodeInspectus demo

Reproduce the V2.1 proof: the codeinspectus@2.1.0 package scans an immutable public Rich commit, finds one high-confidence GitHub Actions expression-injection pattern, applies GitHub's documented intermediate-env remediation in a temporary clone, confirms it as 1 resolved, 0 remaining, 0 introduced, 0 not rechecked, then creates and verifies sealed evidence for both states. Run the reproduction script or read the scanner-derived case study. The recorded pre-publication run used the exact V2.1 tarball; the script defaults to npm after publication. The case uses the ai scanner class to isolate stable native behavior; use a normal full scan for broad repository coverage.

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 Pub — first-party, exact-version Dart/Flutter dependency matching and CycloneDX/SPDX inventory from pubspec.lock, backed by a bundled offline OSV Pub snapshot
• CodeInspectus native checks — client-side secret/bundle exposure, Supabase RLS / inverted-auth (the CVE-2025-48757 class), prompt-injection sinks, model-produced tool arguments reaching Node, Python, or narrowly supported Go, Java, C#, PHP, Rust, and Ruby shell sinks without a visible guard, general model output reaching JavaScript eval/Function or import-proven shell-string APIs, conventional Next.js and import-proven Express admin API handlers missing visible authentication or server-side authorization, Supabase Edge Functions with explicit anonymous deployment but no request authentication or privileged-operation authorization, client-writable user_metadata authorization, and unsanitized model/user output rendered via dangerouslySetInnerHTML (XSS / LLM05), plus explicit API-boundary leaks, raw request-to-database writes, sensitive logging, and evidence-gated security-header/CSP/Referrer-Policy/Permissions-Policy/session-cookie/Supabase-CAPTCHA configuration checks. Separate first-party packs cover six narrow Flutter/Dart source failure modes and eight bounded Android/iOS repository-configuration failures, plus four React Native and two Expo framework-specific mobile failures. A bounded Python AI/API pack covers ten narrow Django, Flask, FastAPI, Starlette, Jinja, OpenAI, Anthropic, LangChain, and OS-command source failures. Separate Go, Java, and C# AI packs each contribute one exact official OpenAI SDK tool-argument-to-shell rule; the PHP pack contributes one equivalent rule for the community-maintained openai-php/client ecosystem, and the Rust pack contributes one bounded rule for the community-maintained async-openai ecosystem. A Ruby pack contributes one equivalent rule for the exact official openai gem. A Firebase configuration pack contributes three literal public-write rules for Firestore, Cloud Storage, and Realtime Database. A GitHub Actions pack contributes two workflow rules for direct untrusted-context shell interpolation and exact pull_request_target checkout-and-execute chains.

The shipped manifest contains 94 curated detections: 72 first-party native rule IDs (29 JavaScript/TypeScript, 6 Flutter/Dart, 4 Android, 4 iOS, 4 React Native, and 2 Expo, plus 10 Python AI/API, 1 Go AI, 1 Java AI, 1 C# AI, 1 PHP AI, 1 Rust AI, 1 Ruby AI, 3 Firebase configuration, 2 GitHub Actions workflow, and 2 JavaScript baseline SAST rules), 18 Opengrep-owned SAST rules, and 4 custom Gitleaks rules. All 20 Opengrep YAML rules remain physically active: the two native-owned rules reconcile exact results and fall back to Opengrep on mismatch or native unavailability. Opengrep, Gitleaks, and Trivy are optional, managed, additive engines.

CodeInspectus explains each engine's coverage, license, platform-specific size, and required action before asking permission. After approval it downloads the official, SHA-pinned engine binaries, stores them outside the npm package, 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 risky AI/vector-store 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

Prerequisite: Node.js ≥22. Node 24 LTS is recommended. No separate engine or Cosign installation is required. CodeInspectus can bootstrap a SHA-pinned Cosign verifier inside ~/.codeinspectus/ after approval; signature verification remains fail-closed. On Linux, the current upstream Opengrep assets require glibc. Alpine/musl remains supported for native CodeInspectus rules, Gitleaks, and Trivy, but setup marks Opengrep unavailable before any download and reports aggregate scan coverage as partial when it is selected.

The official @contentauth/c2pa-node validator is an optional peer because its upstream package downloads a platform-native binding during its own lifecycle script. Normal CodeInspectus installs do not install that peer. Install it explicitly alongside CodeInspectus when local C2PA validation is required; otherwise candidate assets report partial content_provenance coverage.

# Interactive: inspect coverage, licenses, and sizes; then approve all or choose components.
npx codeinspectus setup

# Automation after an operator has reviewed the plan:
npx codeinspectus setup --status
npx codeinspectus setup --all
npx codeinspectus setup --select opengrep,gitleaks

On a terminal, a first bare npx codeinspectus run opens this guided setup. MCP clients continue to start over piped stdio and expose codeinspectus_setup: agents must request a plan, show it, ask permission, then call install with confirm_downloads=true. Declined choices are saved so users are not repeatedly prompted; setup --reset clears them.

Setup first checks local state without network access. It downloads only missing, mismatched, or newly pinned binaries, verifies them against the immutable lockfile shipped in the npm package, and atomically installs them under ~/.codeinspectus/. It refreshes the offline Trivy vulnerability DB only when it is missing, lacks rescan provenance, or is more than seven days old. Rule-only CodeInspectus upgrades therefore download nothing. After setup, scans perform zero network I/O.

Every scan and codeinspectus_list_rules response includes structured engine_setup state: ready, repair_required, db_refresh_recommended, or unsupported_platform. MCP agents are instructed to explain non-ready state and obtain approval through codeinspectus_setup. There is no silent npm postinstall download. repair-engines remains available for advanced/manual use; the older install-engines command remains a compatibility alias.

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 repair-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

CLI, CI, and local evidence workflows

codeinspectus scan . --format sarif --output results.sarif
codeinspectus scan . --format csv --output findings.csv
codeinspectus scan . --format sarif --output results.sarif --fail-on-severity high
codeinspectus scan . --baseline SCAN_ID --fail-on-new-severity high
codeinspectus scan . --diff origin/main --head HEAD
codeinspectus scan . --working-tree --base HEAD
codeinspectus bundle create SCAN_ID --output-dir /outside/repository/scan-results
codeinspectus bundle verify /outside/repository/scan-results
codeinspectus bulk scan /absolute/path/to/local-repositories --concurrency 2
codeinspectus history scan . --from BASE_SHA --to HEAD_SHA --since 2026-07-01 --until 2026-07-30 --max-commits 20
codeinspectus issue export SCAN_ID CI-0001 --adapter github --visibility private

Git-scoped scans retain full repository context, tag changed versus supporting-context findings, and report exact resolved revisions and explicit completeness limits. They never checkout, reset, stage, or modify the repository. See docs/GIT-SCOPED-SCANS.md.

Sealed bundles retain redacted JSON, SARIF, Markdown, coverage, provenance and an additive canonical scan record with content hashes for every artifact. Verification is mandatory before bundle export or comparison. See docs/SEALED-SCAN-BUNDLES.md.

CSV is a deterministic spreadsheet-safe projection of the canonical JSON model. It always retains an explicit scan/coverage row, even with zero findings, and neutralizes formula-triggering cells. See docs/CSV-EXPORT.md for the stable column contract.

The V2 TypeScript SDK is available from codeinspectus/sdk. It is a bounded, shell-free wrapper around the exact installed local CLI and exports versioned finding, coverage, history, baseline, triage and bundle types without duplicating scanner logic. See docs/TYPESCRIPT-SDK.md.

Bulk mode scans already-existing repositories under one explicit local parent with bounded concurrency, per-repository isolation and an atomic resumable manifest. It never clones or requires a GitHub account. See docs/BULK-SCANNING.md.

Repository-history mode is separately opt-in and requires exact revision, UTC date and commit-count bounds. It scans isolated immutable snapshots, marks shallow or truncated history partial, and never describes an old finding as current or a historical secret as active. See docs/REPOSITORY-HISTORY.md.

Issue adapters generate one redacted, review-required GitHub, Jira or Linear JSON payload without authentication or submission. Destination visibility is mandatory and public/private disclosure warnings remain in the artifact. See docs/ISSUE-PAYLOADS.md.

The first command is report-only: findings are retained but do not fail complete scans. The second enforces a severity threshold. Both fail closed with exit 2 when aggregate coverage is partial or unknown; coverage takes precedence over finding severity. Exit 1 is reserved for threshold findings after complete coverage. See the CLI command reference and CI policy and SHA-pinned GitHub Actions workflow for the full exit contract, SARIF upload, artifact privacy, and fork/Dependabot behavior.

Baseline enforcement fails only on findings proven new against a compatible explicit stored scan. Incompatible, partial, or unknown comparison evidence fails closed with exit 2. Local triage adds append-only review context without hiding or changing findings. See baselines and local triage for commands, exact matching, bounds, storage, redaction, schemas, and audit behavior.

Local scan history can be listed, inspected, rerun, and compared without network access:

codeinspectus scans list --repository "$PWD"
codeinspectus scans show SCAN_ID
codeinspectus scans rerun SCAN_ID
codeinspectus scans compare OLD_SCAN_ID NEW_SCAN_ID --format json

Comparison is evidence-gated: absence is Resolved only after compatible producer components and complete like-for-like coverage; otherwise it is Not rechecked / unknown. See the scan history and comparison contract for filters, bounds, V1.x compatibility, corruption handling, and Reopened provenance.

The shipped agent rules also define an approval-gated one-finding remediation workflow: investigate one exact finding, propose the regression and smallest patch, edit only after separate approval, test, then rescan against the exact prior scan. Only a CodeInspectus resolved result supports a scanner-resolution claim; not_rechecked remains an explicit proof gap.

Two explicit, optional workflows extend investigation without changing scanner truth:

• Threat-model and knowledge-base review treats every repository-controlled document as untrusted context. Documents can explain or prioritize raw findings, but never suppress, downgrade, override, or mark them resolved.
• Bounded multi-agent review keeps deterministic findings and agent interpretations in separate evidence lanes, applies explicit agent/time/scope/cost limits, and requires an exact-prior deterministic rescan before any scanner-resolution claim.

These rules are included in the npm package under agent-rules/; normal CLI and MCP scans do not load documents, invoke models, or depend on either workflow.

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.

Source Integrity — V3.1

V3.1 activates the first capability in the stable, non-CWE repository_trust contract across scan, rescan, CLI JSON, MCP structured output, SARIF metadata, and the TypeScript SDK. It deterministically inspects supported repository text for bidirectional overrides/unbalanced controls, zero-width or default-ignorable token characters, concealed Unicode tag sequences, encoded variation-selector runs, and a bounded set of mixed-script identifier confusables.

Every artifact reports escaped code points, Unicode names, exact file/line/code-point column, UTF-8 byte offset, context classification, validator identity, confidence, limitations and an approval-required proposed action. Very long tag/variation runs retain the exact span and sequence length while capping rendered evidence to 64 code points; dense candidates and identifiers are bounded before output materialization and incomplete work is reported as partial. Initial BOMs, legitimate RTL text, emoji variation/ZWJ/tag sequences, international-language joiners and ambiguous confusables are suppressed or kept non-destructive. Scans never edit files; cleanup requires explicit approval for the named file and marker, a reversible edit by the user's coding agent, tests, and a rescan.

This is source-integrity protection, not AI-authorship detection. V3.2's explicit-attribution and C2PA results are separate capability records; hidden Unicode is never promoted into either. CodeInspectus does not claim that hidden Unicode is a Claude watermark or evidence of AI generation.

AI Provenance Audit — V3.2

V3.2 activates two more read-only repository_trust capabilities:

• explicit_ai_attribution deterministically inspects bounded source/config headers, local git commit trailers, and supported image EXIF/XMP/IPTC/PNG metadata for explicit AI-generator or AI-source declarations. These are verified observations of declarative records—not proof that the statement is truthful, proof of vendor origin, or a statistical authorship judgment.
• content_provenance uses the optional official Content Authenticity Initiative Node validator to inspect supported local C2PA Content Credentials. Results distinguish valid, trusted, invalid, present-but-inconclusive, and remote-reference evidence. Offline scans never fetch a remote manifest, OCSP response, trust list, or revocation endpoint.

Every parser and traversal path is bounded, rejects symbolic links, and reports exclusions as partial. C2PA and legal/licensing attribution records are protected evidence and are never cleanup-eligible. V3.2 does not edit files, remove metadata, rewrite text, inspect pixels/audio/video frames, or claim that absence of a marker means human authorship. Statistical text-watermark verification remains unavailable until an authoritative, independently verifiable detector with calibrated operating thresholds is available.

See the V3 migration guide for schema and SDK changes.

detected_technologies explains the bounded repository signals CodeInspectus saw. pack_coverage separately reports how many registered native analyzers and rules actually ran. A pack state of ran means those listed rules executed; it is not a claim of complete security coverage for the named language or framework. not_applicable means the installed pack did not match detected project technology; not_run, partial, and unavailable distinguish scanner filtering from incomplete or failed execution. The Flutter pack runs only when bounded repository signals identify Flutter; an ordinary Dart package does not activate it. Android and iOS packs likewise require bounded platform-project evidence and report their own platform metadata. React Native and Expo are separate technology-gated packs: an Expo project can run both, while a bare React Native project never implies that Expo configuration rules ran. The Python AI/API pack requires bounded Python/package/framework evidence. It reports unsupported syntax, source bounds, and deliberately excluded corpora as explicit coverage notes instead of inferring that omitted source is safe. The Go, Java, and C# AI packs require their language plus the exact official OpenAI SDK dependency. The PHP pack requires PHP plus exact openai-php/client or openai-php/laravel Composer evidence; those packages are community-maintained, not official OpenAI SDKs. A generic openai framework tag alone cannot activate it. The Rust pack requires Rust plus exact async-openai Cargo dependency evidence; that crate is community-maintained, not an official OpenAI SDK. The Ruby pack requires Ruby plus exact official openai production Gemfile or runtime gemspec evidence; lockfiles remain Ruby-language evidence because they do not preserve dependency groups. A generic openai framework tag alone cannot activate Go, Java, C#, PHP, Rust, Ruby, or Python analyzers. dependency_coverage separately reports whether the native Pub matcher ran, which lockfiles and eligible packages it analyzed, what it deliberately skipped, and the bundled snapshot version.

Honest claims (please read)

• "No egress" is precise: zero egress at scan time. Engine binaries and the Trivy DB are fetched only by explicit setup/repair commands from verified sources, with SHA256 verification. Normal CodeInspectus installation does not install the optional C2PA peer; users who explicitly install that upstream package allow its lifecycle download. Nothing is downloaded during a scan. 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.
• Flutter/Dart native coverage is six narrow, first-party structural checks: ci-flutter-tls-verification-disabled, ci-flutter-sensitive-shared-preferences, ci-flutter-webview-untrusted-content, ci-flutter-sensitive-log, ci-flutter-supabase-privileged-key-client, and ci-flutter-cleartext-network. They inspect Dart tokens and local source structure without type resolution or whole-program dataflow. Project-root scans exclude generated files and test/example corpora unless those paths are scanned directly. Unreadable Dart files, files over 2 MiB, and source beyond the 10,000-file/64 MiB project bounds are skipped and named in pack_coverage. This is repository evidence, not runtime mobile testing or complete Flutter coverage. Native Pub vulnerability/SBOM analysis is a separate vuln-class engine, so it can run for Flutter and plain Dart projects without implying that the Flutter source pack applied. Android/iOS configuration is handled by separate native platform packs; existing Trivy behavior is unchanged. The shipped Flutter TP/FP/fixed corpus locks exactly one expected finding per rule in the vulnerable project and zero findings in both the near-miss and remediated projects. E23/E24 exercise that corpus through the built MCP stdio server, including pack execution accounting and redaction of the planted synthetic sentinel.
• Android/iOS native coverage is eight narrow, first-party repository-configuration checks. Android flags an explicitly debuggable release, effective production cleartext opt-in, production trust of user-added CAs, and an exported AndroidX FileProvider. iOS flags global ATS arbitrary loads, insecure production-domain exceptions, weak production-domain TLS policy, and disabled default data protection. The packs parse bounded, literal XML/Xcode repository evidence; they never run Gradle, Xcode, or target code, and parsed configuration is never executed. Dynamic build settings, arbitrary Android product flavors, full manifest merging, provisioning profiles, runtime behavior, and complete mobile-platform security remain out of scope. Skipped or unsupported configuration is named in pack_coverage, not inferred as secure or vulnerable. The shipped Android/iOS TP/FP/fixed corpus and E25/E26 lock exact findings, zero-finding near-miss/remediated states, provenance, redaction, execution accounting, and bidirectional same-path rescan behavior.
• React Native/Expo native coverage is six narrow, first-party structural checks. The React Native pack flags sensitive credential writes to proven AsyncStorage receivers, untrusted route/ deep-link content entering an imported JavaScript-enabled WebView, explicit mixed-content opt-in, and universal-origin access from a file-backed WebView. The separate Expo pack flags sensitive server environment values placed in public app config and unsigned updates fetched over cleartext production URLs. Both use bounded read-only parsing and never import, execute, or evaluate target code/configuration. Dynamic values, spreads, unresolved guards, unreadable/oversized input, and bounded-out source become named coverage limitations rather than inferred findings. Framework evidence is required; ordinary JavaScript/TypeScript or native Android/iOS skeletons do not activate these packs. This is intrafile/static repository evidence, not complete dataflow, deployment proof, or runtime mobile testing. The shipped React Native/Expo TP/FP/fixed corpus and E30/E31 lock exact findings, precision, redaction, provenance, execution accounting, and same-path resolution/reintroduction.
• Python AI/API native coverage is ten narrow first-party structural checks. The pack flags hardcoded framework signing secrets, credentialed all-origin CORS, request-controlled file responses and redirects, request-controlled template source, and unsanitized OpenAI/Anthropic output returned as HTML, explicit LangChain FAISS pickle-deserialization opt-in, and request-controlled complete URLs fetched by a proven LangChain WebBaseLoader, plus medium-confidence potential prompt-injection sinks where request input reaches proven OpenAI/Anthropic privileged instructions or shares the LLM call with configured tool access, and model tool arguments reaching proven Python shell-execution APIs without a visible checked guard. It uses a bounded Lezer syntax gate plus source-ordered intrafile analysis; it never imports or executes target Python. There is no type checker, module graph, interprocedural flow, or path-sensitive branch merge. Lezer-validated format strings are retained as opaque dynamic values, so their replacement expressions are not inspected; leading-tab indentation still fails closed with a named coverage note. Generated, migration, dependency, build, test, fixture, demo, sample, and example trees are excluded from project-root scans. Unsupported, malformed, symlinked, unreadable, oversized, or bounded-out source is reported, not inferred as secure. The shipped Python AI/API TP/FP/fixed corpus and E32/E33 lock exact findings, precision, redaction, provenance, execution accounting, and same-path resolution/reintroduction.
• Go AI native coverage is one narrow, medium-confidence structural check. ci-go-llm-tool-argument-command-execution requires the exact official OpenAI Go SDK Chat Completions tool-call argument shape and import-proven os/exec invocation of a recognized shell with its command flag. It supports direct aliases, encoding/json.Unmarshal, one local JSON parser, and one local command wrapper. Checked rejection/approval or allowlist guards and validated replacement values suppress findings. It does not cover general Go SAST, other model SDKs, cross-module dispatch, runtime sandboxing, or complete agent safety. The shipped Go TP/FP/fixed corpus and E37/E38 lock exact findings, zero-finding safe and fixed states, provenance, pack dispatch, and same-path resolution/reintroduction.
• Java and C# AI native coverage is one narrow, medium-confidence structural check per ecosystem. Each requires the exact official OpenAI SDK dependency and tool-call argument shape before model data can reach a recognized, actually-started shell process. Direct aliases, one local parser, and one local command wrapper are supported; checked rejection/approval or allowlist guards and validated replacement values suppress findings. These packs do not provide general Java or C# SAST, cross-module dispatch, runtime sandbox proof, or complete agent review. The shipped Java and C# TP/FP/fixed corpora and E39-E42 lock exact findings, safe/fixed silence, provenance, language-gated dispatch, and same-path resolution/reintroduction.
• PHP AI native coverage is one narrow, medium-confidence structural check. ci-php-llm-tool-argument-command-execution requires exact community-maintained openai-php/client or openai-php/laravel Composer evidence and model tool-call function->arguments reaching exec, system, shell_exec, or passthru. It supports direct aliases, associative json_decode, one local parser, one local command wrapper, and one exact mapped variadic method dispatch. Checked approval/full-command allowlists and validated replacement values suppress findings; first-token executable checks do not neutralize shell metacharacters and do not suppress them. It does not provide general PHP SAST, cross-file flow, runtime sandbox proof, or complete agent review. The shipped PHP TP/FP/fixed corpus and E43/E44 lock exact findings, safe/fixed silence, provenance, language-gated dispatch, and same-path resolution/reintroduction.
• Rust AI native coverage is one narrow, medium-confidence structural check. ci-rust-llm-tool-argument-command-execution requires exact community-maintained async-openai Cargo evidence and recognized model tool arguments reaching an import-proven standard/Tokio process shell or literal Bollard Docker exec shell vector. It supports direct aliases, serde_json extraction, one recognized generate_function_call result, and one local command wrapper. Checked approval/allowlist rejection and validated replacement values suppress findings. It does not provide general Rust SAST, rustc/type/Cargo-graph resolution, cross-crate flow, runtime container/sandbox proof, or complete agent review. The shipped Rust TP/FP/fixed corpus and E45/E46 lock exact findings, safe/fixed silence, provenance, language-gated dispatch, and same-path resolution/reintroduction. async-openai is community maintained and is not represented as an official OpenAI SDK.
• Ruby AI native coverage is one narrow, medium-confidence structural check. ci-ruby-llm-tool-argument-command-execution requires exact official production openai Gemfile or runtime gemspec evidence and recognizes Chat tool-call function.arguments or explicitly typed Responses function-tool arguments reaching system, exec, IO.popen, or import-proven Open3 shell execution. It supports direct aliases, JSON.parse command extraction, one local parser, and one local command wrapper. Checked approval/full-command allowlists and validated replacements stay silent. It does not provide general Ruby SAST, Bundler/type resolution, cross-file flow, backtick/percent-x/ spawn coverage, runtime sandbox proof, or complete agent review. The shipped Ruby TP/FP/fixed corpus and E47/E48 lock exact findings, safe/fixed silence, provenance, language-gated dispatch, and same-path resolution/reintroduction.
• Firebase native coverage is three narrow, high-confidence configuration checks. ci-firebase-firestore-public-write, ci-firebase-storage-public-write, and ci-firebase-realtime-database-public-write flag only checked-in literal write grants with no condition or a condition exactly equal to true (Realtime Database boolean/string true). Public reads stay silent because they are often intentional. The bounded parser masks Rules comments/strings, requires exact Firestore/Storage service declarations, requires strict JSON for Realtime Database, excludes non-production/dependency/generated trees, and never runs Firebase tooling or target code. It does not evaluate helper functions, deployed policy, IAM, App Check, or runtime access. The shipped Firebase TP/FP/fixed corpus and E49/E50 lock exact findings, safe/fixed silence, provenance, pack dispatch, and same-path resolution/reintroduction.
• GitHub Actions native coverage is two narrow workflow checks. ci-github-actions-untrusted-expression-command flags direct documented attacker-controlled github expressions embedded in run; using an intermediate env value or action with input stays silent. ci-github-actions-pwn-request requires the complete high-risk chain: exact pull_request_target, checkout of the pull request's untrusted ref into the default workspace, and subsequent execution of checked-out code or a local action. Checkout v7 is treated as protected unless allow-unsafe-pr-checkout is explicitly enabled. The bounded YAML 1.2 parser reads only direct .github/workflows/*.yml/.yaml files and never executes workflows or target code. It does not provide general workflow taint analysis, custom-action analysis, artifact-flow analysis, permission-policy proof, or runner/runtime verification. The shipped GitHub Actions TP/FP/fixed corpus and E51/E52 lock exact findings, safe/fixed silence, provenance, dispatch, and same-path resolution/reintroduction.
• 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 JWT or modern sb_secret_ key value in client-reachable code (critical), and a privileged key behind a client-exposed env prefix such as NEXT_PUBLIC_… (high). Exact-format sb_publishable_ keys remain silent because they are public by design. Each secret class is bounded to 1,024 remediation locations per file; repeated values on one minified line collapse, separate source lines remain visible, and reaching the bound emits an explicit coverage note.
• 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 follows one local function-component hop when a destructured prop is passed directly to __html; sanitizing at the JSX call site stays silent. It does not trace cross-file or arrow-component props, database rows, template data, object-spread sinks, or custom sanitizer wrappers.
• Model output passed to dynamic execution is flagged (ci-ai-llm-output-dynamic-execution). CodeInspectus follows direct and split-variable intrafile flows from recognized OpenAI, Anthropic, Google GenAI, Vercel AI SDK, Cohere, Groq, and Mistral call/output shapes into global eval/Function, import-proven Node child_process.exec/execSync, or import-proven Execa command-string APIs. The rule is high severity and medium confidence (CWE-94/78/1426; OWASP LLM05). Constants, fixed allowlisted dispatch, execFile/spawn argument arrays, shadowed globals, and explicitly validated replacement values stay silent. Cross-file flows, custom model wrappers, stream accumulation, indirect sink aliases, and runtime sandbox or approval state are not resolved.
• Conventional Next.js admin API routes need authentication and authorization (ci-ai-nextjs-admin-route-no-authz). Pages Router handlers under pages/api/admin and App Router handlers under app/api/admin/**/route.* are flagged when either boundary is not visible in the file (high, medium confidence; CWE-862/863/306; OWASP A01 and API5). Recognized server-session/token checks plus role/permission decisions stay silent; Supabase client-writable user_metadata is deliberately not accepted as authorization. Non-admin/public routes stay silent. The bounded detector follows exact direct, one-local-helper-hop evidence; unsupported cross-file middleware, custom guard semantics, and dynamic dispatch remain explicit unknowns.
• Import-proven Express admin routes and explicitly anonymous Supabase Edge Functions are checked. Express analysis requires a literal admin route and direct handler evidence, respects supported middleware/mount ordering, and reports unresolved custom middleware as coverage metadata. Edge analysis treats default or explicit verify_jwt = true as platform authentication; only explicit anonymous deployment requires visible request authentication, and service-role/admin operations additionally require visible server authorization. Cross-module custom wrappers remain unknown.
• Server/API-boundary checks are narrow and code-visible. Four JavaScript/TypeScript analyzers flag client-visible raw/internal error details (CWE-209), explicit credential fields in response objects (CWE-201), whole request objects passed directly to common Prisma/Supabase/Mongoose writes without visible validation or allow-listing (CWE-915), and explicit credentials/headers/cookies or auth/payment request bodies sent to logs (CWE-532). They use intrafile dataflow and prefer silence when validation, a public-error mapper, or an explicit field projection is visible. Build output and minified vendor files are outside these source checks. They do not claim generic response minimization, complete business-authorization review, or runtime/gateway verification.
• Runtime security controls use three evidence states, never absence-as-vulnerability. Supported header/CSP/cookie/CAPTCHA controls report verified_in_repository, insecure_configuration_found, or not_verifiable_from_repository. Only an explicit insecure repository configuration becomes a finding; missing headers, dashboard-only CAPTCHA, dynamic/conflicting layers, and hosted/gateway configuration remain metadata with no posture penalty. “Verified” means only that recognized source configuration passed the rule’s narrow literal check; it is not runtime or complete policy proof. Current findings cover security headers explicitly disabled or neutralized, production CSP with bare wildcard or 'unsafe-eval' script sources, an effective literal Referrer-Policy: unsafe-url, unrestricted camera/microphone/geolocation delegation in a literal Permissions-Policy, auth/session cookies with explicit insecure attributes, and checked-in Supabase CAPTCHA enablement paired with a recognized signup, password/OTP/SSO/Web3 signin, or password-reset call missing captchaToken. The CAPTCHA finding describes an integration failure that Supabase should reject—not a bot-protection bypass. CodeInspectus still does not prove deployed headers, gateway rate limits, DNS mail posture, complete CSP quality, runtime overrides, or behavioral authentication. SPF/DMARC absence is not emitted as a repository finding because authoritative state is live DNS, and DMARC p=none is a valid monitoring mode; verify mail domains explicitly at deployment time.

Language support

Plainly, what runs on what. The commodity engines are broad; the CodeInspectus native checks are predominantly JavaScript/TypeScript, plus targeted Flutter/Dart, Android/iOS configuration, React Native, Expo, Python AI/API, and one-rule Go, Java, C#, PHP, Rust, and Ruby AI packs, plus Firebase Security Rules and GitHub Actions workflow configuration packs. A Python repository gets native coverage only for the ten documented framework/source shapes; a Go/OpenAI, Java/OpenAI, C#/OpenAI, PHP/openai-php, Rust/async-openai, or Ruby/OpenAI repository gets only its one documented tool-execution rule. Other unsupported native-pack ecosystems still receive the selected commodity-engine coverage but no native pack applies. Plain Dart without Flutter also does not activate the Flutter source pack (the separate native Pub SCA/SBOM engine can still run). This is stated so you don't infer broader coverage than the executed pack reports.

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.

Relevant findings can also carry OWASP API Security Top 10 (2023) category tags. Those tags describe the detected failure pattern; they are not an OWASP API review, coverage score, compliance claim, or certification.

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 | native Pub] + applicable native packs
      → normalize → dedup (incl. Trivy/native-Pub advisory aliases + secret overlap)
      → CWE-keyed findings + runtime-control evidence → compliance map → compact JSON + summary
ALL LOCAL. NO NETWORK EGRESS AT SCAN TIME.

Example reports

• V2.1 public-repository scan → fix → rescan → sealed-evidence case study
• V1.5 public-repository scan → fix → rescan case study
• 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       # 53 MCP stdio evals across the shipped verification fixtures
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/**, src/packs/**). 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

• 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 is licensed under Apache-2.0. The separately downloaded engines retain their own licenses: Opengrep (LGPL-2.1), Gitleaks (MIT), and Trivy (Apache-2.0). Engine binaries are not included in the npm package. See THIRD-PARTY-NOTICES.md for versions, copyright notices, upstream source, and license links. The transformed Pub advisory snapshot is derived from OSV.dev records sourced from the GitHub Advisory Database and remains CC BY 4.0; attribution and transformation details ship beside the snapshot in detection-db/osv-pub/LICENSE-PROVENANCE.md.