code-audit-mcp

Audit your whole stack with a single question to your agent.

An MCP server (stdio) that gives Claude Code (and any MCP-compatible agent) a complete audit of your project: dead code, dependencies, security, CVEs, complexity and lint. It covers Laravel/PHP, Python (FastAPI, Flask), TypeScript/React/Next.js and Flutter/Dart, including monorepos (backend/ + frontend/, Laravel with React, a Flutter app next to its backend).

"Audit this repo" -> within a minute you get a report prioritized by severity, with false positives already flagged and the next step for every finding.

Why it is different

• Zero installation in the project. vendor/, node_modules, the venv and .dart_tool stay untouched. Tools run through uvx/npx or from ~/.cache/code-audit-mcp, and the first run downloads them on its own.

• Detects stacks by itself. It walks the repo, finds every project (even when several coexist) and audits each one with its native tools.

• Built to be used by an LLM. Findings are normalized (file, line, rule, severity, message and how to fix it), sorted by severity and free of the noise of ten different tools.

• Flags false positives and explains why. Routes registered by decorator, Laravel and Livewire hooks, Pydantic fields or the HEAD strings that bandit mistakes for passwords do not count in the summary. This way the agent never proposes deleting code that is in fact used.

• Checks git history, not just the current code. A secret committed a year ago and deleted later is still compromised, and it finds it.

• Secrets never appear in the output. They are used to group findings and to tell whether they are still in the code, but the value never leaves the server.

Related MCP server: ARGUS

Two tools

Checks: dead_code · dependencies · security · vulnerabilities · complexity · lint

Context savings. Responses are a single compact JSON (no structuredContent copy and no indentation). Messages that repeat identically across several findings of a rule are emitted only once in rules. check returns 50 findings per engine (limit) and leaves out likely false positives (false_positives=true to see them). Narrow results with severity (minimum) and file (glob or prefix, e.g. app/Http/**) and paginate with offset; the response tells you in next how to continue.

What it covers in each stack

In every stack, security adds gitleaks over the whole git history and sensitive files that are tracked, not ignored, or deleted but still in history.

Flutter/Dart. dart analyze needs .dart_tool/ (flutter pub get); without it, it is skipped, because generating it would write to the project. The dart from the Flutter SDK is used if available. The Pub ecosystem has few advisories published in OSV: 0 CVEs does not guarantee there are none.

Security: what generic scanners miss

On top of the public semgrep packs, it ships its own rules (code_audit_mcp/rules/stack.yml) for the problems of these stacks:

• Postgres/SQL: SQL with interpolation or concatenation in DB::raw/select, whereRaw/orderByRaw…, cursor.execute(f"…"), text(f"…"), pool.query(\…${x}`), $queryRawUnsafe`; connection URLs with a password in the code.

• Redis: KEYS (blocks the server; use SCAN) and FLUSHALL/FLUSHDB in application code.

• Laravel: env() outside config/ (breaks with config:cache), $guarded = [], {!! !!} in Blade, eval.

• FastAPI/Flask: debug=True, CORS * with credentials, SECRET_KEY in the code.

• React: dangerouslySetInnerHTML, tokens in localStorage.

• Flutter/Android/iOS: badCertificateCallback accepting any certificate, tokens and passwords in SharedPreferences, usesCleartextTraffic/debuggable in the release AndroidManifest.xml and NSAllowsArbitraryLoads in Info.plist.

Secrets and sensitive files:

• Files: .env*, private keys and certificates (.pem, .key, .p8, id_rsa, .p12…), credentials (.pypirc, .netrc, .pgpass, .git-credentials, .htpasswd, Android key.properties; .npmrc/.yarnrc.yml and Composer's auth.json only if they carry tokens), data dumps (.sqlite, .db, *dump*.sql, .sql.gz, .bak) and logs. They are reported when tracked in git, one git add . away from being committed (not ignored) and deleted but still present in history, with commit, author and date so you know what to rotate.

• Secrets in history (gitleaks): tokens and keys that were ever committed on branches, tags or remotes, even if they are no longer in the code. One finding per secret, with the oldest commit and whether it is still in the code.

False positives

Dead-code detectors cannot see implicit usages. Every likely false positive of this kind comes out with likely_false_positive and the reason, and does not count in the severity summary:

• Python: functions registered by decorator (@app.get, @router.post, @bp.route, @command…), Pydantic/SQLAlchemy/dataclass/Enum model fields, Alembic migrations, framework hooks. In dependencies, packages that are used without being imported (asyncpg/psycopg through the connection URL, python-multipart, uvicorn, alembic…).

• Laravel: Laravel/Livewire/Filament hooks, Eloquent accessors and scopes, public members of Livewire/Filament components and names referenced from Blade views.

• knip: laravel-vite-plugin entries (passed to knip as entry) and files of plugins that could not be loaded.

• bandit: import subprocess (B404), commands with literal arguments (["git", "status"], B603/B607) and "passwords" that are not (HEAD, refresh_token:, uppercase constants; B105-B107).

Tests count as usage, but their findings are not reported.

Optional configuration: .code-audit.toml at the project root

exclude = ["legacy/**"]                         # paths to skip
ignore = ["B404", "DEP002:ruff", "unused-method:*Resource::form"]  # rule or rule:name (globs)

[dead_code]
ignore_names = ["on_*"]
ignore_decorators = ["@command"]   # Python
min_confidence = 60                # Python (vulture)

[complexity]
threshold = 11                     # minimum cyclomatic complexity (C = 11-20, D = 21-30…)

[security]
registry = true                    # public semgrep packs (needs network)

Requirements

• uv (always).

• PHP 8.2+ and Composer for PHP projects; Node.js for JS/TS projects; bun if the project uses bun.lock; Flutter or Dart SDK for linting Dart projects.

• gitleaks downloads itself (official GitHub binary, sha256 verified) to ~/.cache/code-audit-mcp/bin, unless it is already in the PATH. CVEs for yarn.lock and pubspec.lock are queried against the OSV.dev API (network).

• For complete results, the project must have its dependencies installed (composer install, npm ci, venv). Without them, each affected engine says so in skipped or notes.

The first run downloads the tools (a few minutes). On a large Laravel project, dead code takes 1-2 minutes; the rest of the checks, seconds.

Installation

Globally in Claude Code:

claude mcp add -s user code-audit /path/to/code-audit-mcp/bin/code-audit-mcp-server

Per project (.mcp.json):

{
  "mcpServers": {
    "code-audit": {
      "command": "/path/to/code-audit-mcp/bin/code-audit-mcp-server",
      "args": []
    }
  }
}

Environment variables

Development

uv sync
uv run pytest
uvx ruff check .

Layout: core.py (processes, cache, finding format), detect.py (stacks), config.py, engines/{python,php,node,dart,shared}.py, checks.py (which engine resolves each check), server.py. The own semgrep rules are tested against tests/fixtures/rules, where every line that must be detected carries BAD:<rule>.

Contributing

Issues and pull requests are welcome; see CONTRIBUTING.md. Security reports go through SECURITY.md. Release notes live in CHANGELOG.md and planned work in ROADMAP.md. Licensed under the MIT License.

code-audit-mcp: the quality and security report your agent can read, prioritize and fix.