mcp-audit-scanner 0.4.0

Privacy-first security scanner for MCP server configurations

mcp-audit

Privacy-first security scanner for MCP server configurations.

Free & open source. Apache 2.0, all features included — no paid tier, no license keys required.

Privacy-first: mcp-audit collects no telemetry. Every scan runs locally. See docs/telemetry.md for the full policy.

MCP (Model Context Protocol) servers give AI agents access to your tools, files, APIs, and databases. Misconfigured or malicious servers can exfiltrate credentials, poison tool behavior, and compromise your development environment — without anything appearing in the UI.

mcp-audit scans your local MCP configurations across all major AI coding clients, connects to running servers to inspect what agents actually see, and flags security issues across individual servers and dangerous cross-server combinations.

Features

• Auto-discovers MCP configs across 8 clients (Claude Desktop, Cursor, VS Code, Windsurf, Claude Code user-level, Claude Code project-level, GitHub Copilot CLI, Augment Code)
• Tool poisoning detection — 11 regex patterns across 5 severity tiers, validated against 6 published exploit PoCs (Invariant Labs, CrowdStrike, CyberArk) with 0% false-positive rate across 22 real servers
• Credential exposure — 9 patterns covering AWS, GitHub, OpenAI, Anthropic, Stripe, Slack, and database URLs
• Transport security — unencrypted connections, elevated privileges (sudo/doas/pkexec/su/run0), wildcard bindings (0.0.0.0, ::), runtime package fetching
• Supply chain — typosquatting detection via Levenshtein distance against 64 known-legitimate MCP servers; SHA-256 hash verification; Sigstore SLSA provenance verification; transitive-dependency CVE lookup via OSV.dev; CycloneDX SBOM generation
• Rug-pull detection — stateful SHA-256 hash comparison of tool descriptions across scans
• Cross-server toxic flows — capability tagging and 7 dangerous pair patterns detecting multi-server attack paths (file-read + network, secrets + network, shell-exec + network, etc.)
• Attack path engine — multi-hop path detection with greedy hitting set algorithm (minimum set of servers to remove to break all attack paths)
• Interactive attack graph dashboard — mcp-audit dashboard opens a D3 force-directed graph in your browser with light/dark mode, click-to-highlight attack paths, and hitting set recommendations
• Live server analysis — connects to running servers via MCP protocol to inspect actual tool definitions
• SAST rule pack — 37 Semgrep rules (28 Python, 9 TypeScript) for MCP server source code
• IDE extension scanner — known-vuln registry, dangerous capability combos, wildcard activation, unknown publisher, sideloaded VSIX, stale AI extensions
• Governance + policy-as-code — YAML governance policies (approved server lists, score thresholds, transport constraints) and custom detection rules; 12 community rules ship bundled and run for every user
• OWASP MCP Top 10 mapping — every finding carries MCP01–MCP10 codes in terminal, JSON, and SARIF (taxonomy block + per-rule relationships); --owasp-report prints a category-level summary; see docs/owasp-mcp-top-10.md
• 5 output formats — terminal (Rich), JSON, SARIF (GitHub Security tab), Nucleus Security FlexConnect, self-contained HTML dashboard
• Continuous monitoring — mcp-audit watch monitors config files in real-time and re-scans on any change
• Fleet deployment — machine-tagged output with --asset-prefix for enterprise-wide aggregation
• Fully offline by default — no data leaves your machine

Enterprise vulnerability management

MCP security findings typically exist in isolation: a developer runs a scanner, sees terminal output, maybe fixes something. Enterprises deploying AI agents across hundreds of developers need the same workflow they use for every other vulnerability class — centralized ingestion, asset correlation, deduplication, ownership, remediation tracking, and SLA reporting.

mcp-audit's --format nucleus output and mcp-audit push-nucleus command align with the Nucleus Security FlexConnect schema — the same ingestion pipeline that normalizes data from Qualys, Tenable, CrowdStrike, and 200+ other security tools. Validated end-to-end against a live Nucleus instance on 2026-04-23; see docs/nucleus-integration.md.

Tenable WAS has added MCP server detection plugins that scan server-side code for web vulnerabilities, but no other standalone MCP configuration scanner bridges developer-side config analysis (tool poisoning, credential exposure, toxic flows, supply chain risks) with enterprise vulnerability management. Most output to terminal or JSON and stop there.

Free & open source

mcp-audit is released under the Apache License 2.0 and every feature is available to every user. There are no paid tiers, license keys, or gated commands — the full scanner, rule authoring, governance, SAST integration, extension scanning, dashboard, fleet merge, and Nucleus FlexConnect output all ship in the same binary.

---

Install

pip install mcp-audit-scanner

or with uv:

uv add mcp-audit-scanner

Note: The PyPI package name is mcp-audit-scanner (the mcp-audit name was already taken). The CLI command is still mcp-audit. Standalone binaries are also available on GitHub Releases.

For live server connection support:

pip install 'mcp-audit-scanner[mcp]'

For Sigstore-based signature verification (scan --verify-signatures):

pip install 'mcp-audit-scanner[attestation]'

Note: scan --verify-hashes uses stdlib SHA-256 and works without any extra. The [attestation] extra is only needed for Sigstore/Rekor signature verification.

macOS notes

If pip install fails with bad interpreter: /usr/bin/python: no such file or directory, invoke pip through Python directly:

python3 -m pip install mcp-audit-scanner

After installing, if mcp-audit is not found in your shell, add Python's bin directory to your PATH:

echo 'export PATH="/Library/Frameworks/Python.framework/Versions/3.12/bin:$PATH"' >> ~/.zshrc
source ~/.zshrc

If you installed via Homebrew Python, the path will be different — use python3 -c "import sys; print(sys.prefix)" to find it and append /bin.

Quick start

mcp-audit scan                                        # Scan all detected MCP configs
mcp-audit scan --connect                              # Also connect to running servers
mcp-audit scan --format sarif -o results.sarif        # SARIF for GitHub Security
mcp-audit scan --format nucleus -o results.json       # Nucleus FlexConnect output
mcp-audit dashboard                                   # Open interactive attack graph dashboard
mcp-audit dashboard --path demo/configs               # Dashboard against demo data
mcp-audit discover                                    # List detected clients and servers
mcp-audit pin                                         # Lock current state as trusted baseline
mcp-audit diff                                        # Show changes since last pin
mcp-audit watch                                       # Monitor configs and re-scan on changes
mcp-audit push-nucleus --url ... --project-id ...     # Scan and push to a Nucleus project
mcp-audit merge --dir ./scans                         # Merge multi-machine JSON outputs

Supported clients

Client	Config location
Claude Desktop	~/Library/Application Support/Claude/claude_desktop_config.json
Cursor	~/.cursor/mcp.json
VS Code	.vscode/mcp.json (workspace)
Windsurf	~/.codeium/windsurf/mcp_config.json
Claude Code (user)	~/.claude.json
Claude Code (project)	.mcp.json (project root)
GitHub Copilot CLI	~/.copilot/mcp-config.json
Augment Code	~/.augment/settings.json

What it detects

Analyzer	Finding IDs	Examples
Tool poisoning	11 patterns (POISON-001 – POISON-050)	SSH key exfiltration instructions, XML injection markers (<IMPORTANT>), behavioral overrides ("ignore previous instructions"), zero-width Unicode stealth characters
Credential exposure	CRED-001…009	AWS access keys, GitHub tokens, OpenAI/Anthropic API keys, Stripe secrets, database connection strings with embedded passwords
Transport security	TRANSPORT-001…003	Unencrypted remote SSE connections, elevated privilege execution, runtime package fetching via npx/uvx without version pinning
Supply chain	SC-001…003	Typosquatted package names (@modelcontextprotocol/server-filesytem vs server-filesystem), distance-1 substitutions flagged CRITICAL
Rug-pull	RUGPULL-001…003	Tool description changed since last scan (HIGH), new server appeared (INFO), previously tracked server removed (INFO)
Toxic flow	TOXIC-001…007	File-read server + network server (exfiltration path), secret-access server + network server (credential theft), shell-exec server + network server (arbitrary command + exfiltration)

Live server analysis

By default, mcp-audit performs static analysis — it reads config files and inspects the command, args, env vars, and any tool descriptions stored there.

The --connect flag goes further: it connects to each server using the MCP protocol, completes the initialization handshake, and calls list_tools(), list_resources(), and list_prompts() to retrieve the actual definitions the server exposes to the AI agent. Those live definitions are then run through the poisoning analyzer.

This matters because a config file can look completely clean while the server it points to is serving poisoned tool descriptions. Static analysis cannot catch this. Connection-based analysis can.

mcp-audit scan --connect

Requires the optional MCP SDK dependency:

pip install 'mcp-audit-scanner[mcp]'

Connection is best-effort: servers that do not respond within 10 seconds produce an error finding rather than crashing the scan. All static analysis still runs regardless.

Cross-server attack paths

Most MCP security analysis focuses on individual servers. That misses an entire category of risk.

Server A reads files. Server B makes HTTP requests. Neither is malicious alone — they each do exactly what the config says. Together, a prompt injection can instruct the agent to read your SSH keys with A and POST them to an attacker's endpoint with B. No single server ever looked dangerous.

mcp-audit detects 7 categories of these toxic combinations by tagging each server with capability labels (FILE_READ, NETWORK_OUT, SHELL_EXEC, DATABASE, SECRETS, etc.) and checking every server pair for known-dangerous combinations:

ID	Combination	Severity
TOXIC-001	File read + outbound network	HIGH
TOXIC-002	File read + email	HIGH
TOXIC-003	Secret store access + outbound network	CRITICAL
TOXIC-004	File read + shell execution	HIGH
TOXIC-005	Database access + outbound network	HIGH
TOXIC-006	Shell execution + outbound network	CRITICAL
TOXIC-007	Git repository access + outbound network	MEDIUM

† A single server that provides both capabilities of a dangerous pair is also flagged — no second server required.

Attack graph dashboard

mcp-audit dashboard                      # Scan your real MCP environment and open browser
mcp-audit dashboard --path demo/configs  # Use the bundled demo data
mcp-audit dashboard --port 9090          # Custom port
mcp-audit dashboard --connect            # Include live-connection findings

One command runs a full scan, generates a self-contained HTML report, and opens it in your browser. No external dependencies — D3 v7, all scan data, and fonts are embedded inline. No CDN requests are made.

The dashboard shows:

• Force-directed attack graph — your MCP servers arranged around a central AI Agent node. Server nodes are colour-coded by max severity (green = clean, orange = high, red = critical). Toxic flow edges connect pairs with dangerous capability combinations.
• Attack path sidebar — every exploitable multi-hop path listed as a card with severity badge, hop chain, and description. Click a card to highlight the path on the graph with animated dashed lines.
• Hitting set recommendation — at the bottom of the sidebar, the minimum set of servers you can remove to break every attack path. Example: removing fetch alone breaks three separate attack paths.
• Findings table — full findings list with severity filter pills and sortable columns.
• Light/dark mode toggle — pill toggle in the top bar. Preference is applied instantly via CSS custom properties; no page reload required.

The dashboard works against your real MCP environment — whatever mcp-audit scan finds on your machine is what appears in the graph. It is not restricted to demo data.

Rug-pull detection

MCP servers can update their tool definitions at any time. A server can publish clean, trusted descriptions during initial review and silently swap them for malicious ones after developers have granted access.

mcp-audit pin records SHA-256 hashes of every tracked server's configuration as a trusted baseline. Subsequent mcp-audit scan runs compare against that baseline and flag any change as RUGPULL-001 (HIGH).

mcp-audit pin   # Record current state as trusted
mcp-audit diff  # Show what has changed since last pin

Rug-pull state is stored per-config-set at ~/.mcp-audit/state_<hash>.json. All other persistent state (baselines, registry cache, policy, rules, license) uses the platform user-config directory: ~/Library/Application Support/mcp-audit/ on macOS, ~/.config/mcp-audit/ on Linux, %APPDATA%\mcp-audit\ on Windows.

CI/CD usage

mcp-audit exits with code 1 when findings are detected, 0 when clean, and 2 on errors.

# .github/workflows/mcp-security.yml
- name: Scan MCP configs
  run: mcp-audit scan --severity-threshold HIGH

- name: Export SARIF for GitHub Security tab
  run: mcp-audit scan --format sarif -o mcp-audit.sarif

- name: Upload SARIF
  uses: github/codeql-action/upload-sarif@v4
  with:
    sarif_file: mcp-audit.sarif

Where the detection logic comes from

All detection patterns are original implementations based on published security research — no code was copied from existing scanners. Sources include Invariant Labs' tool poisoning disclosure, CrowdStrike's MCP exfiltration research, CyberArk's agent attack demonstrations, the OWASP Agentic Top 10, and MITRE ATLAS agent-specific techniques. Supply chain patterns follow npm package naming conventions; credential patterns follow the publicly documented key formats from AWS, GitHub, OpenAI, Anthropic, Stripe, and others.

1,361 tests validate detection accuracy and guard against regressions.

See PROVENANCE.md for the full list of research sources, framework mappings, and contribution guidelines for new detection rules.

CLI reference

Every command is available to every user — no tier, no license required.

Command	Key flags	Description
mcp-audit scan	--connect, --format, --output, --severity-threshold, --asset-prefix, --baseline, --policy, --verify-hashes, --no-score, --registry, --offline-registry, --rules-dir, --sast, --include-extensions	Run all analyzers and report findings
mcp-audit dashboard	--path, --port, --connect, --no-open	Generate and open the interactive attack graph dashboard
mcp-audit watch	--path, --format, --severity-threshold, --connect	Monitor config files and re-scan on any change
mcp-audit discover	—	List all detected MCP clients and their configured servers
mcp-audit pin	—	Record current server state as a trusted baseline
mcp-audit diff	—	Show configuration changes since the last pin
mcp-audit verify	<package|config-path>	Verify server hashes: pass a package name (@scope/pkg), a config file path, or --all
mcp-audit version	—	Print version string
mcp-audit update-registry	—	Fetch the latest known-server registry from upstream
mcp-audit sast	<path>	Run MCP-aware Semgrep SAST rules on server source code
mcp-audit push-nucleus	--url, --project-id, --api-key, --asset-prefix	Run a scan and push results to a Nucleus Security project via FlexConnect
mcp-audit merge	--dir, --format, --asset-prefix	Merge JSON scan outputs from multiple machines into a fleet report
mcp-audit baseline save [NAME]	--path	Capture a baseline snapshot; NAME is optional (auto-generated if omitted)
mcp-audit baseline list	—	List all saved baselines
mcp-audit baseline compare [NAME]	--path	Compare current config against a saved baseline (defaults to latest)
mcp-audit baseline delete NAME	--yes	Delete a saved baseline
mcp-audit baseline export NAME	--output-file	Write a baseline as raw JSON to stdout or a file
mcp-audit rule validate	<file>	Validate a rule file without running a scan
mcp-audit rule test	<rule> <config>	Test a rule file against a specific MCP config file
mcp-audit rule list	—	List all currently loaded rules (bundled + user-local)
mcp-audit policy validate	<file>	Validate a governance policy YAML file
mcp-audit policy init	—	Scaffold a new governance policy file
mcp-audit policy check	--policy, --result	Check a scan result against a policy file
mcp-audit extensions discover	—	Inventory installed IDE extensions from VS Code/Cursor
mcp-audit extensions scan	—	Analyze installed IDE extensions for security risks

mcp-audit scan flags

Flag	Default	Description
--format	terminal	Output format: terminal, json, sarif, nucleus
--output / --output-file / -o	stdout	File path for json/sarif/nucleus output; parent directories are created automatically
--connect	off	Connect to running servers via MCP protocol
--severity-threshold	INFO	Filter findings and set exit code; exit 1 if any finding at or above this level
--path	auto-detect	Directory to search for MCP configs
--asset-prefix	hostname	Override machine identifier in Nucleus/SARIF output
--no-score	off	Suppress the score/grade panel in terminal output
--registry	bundled	Custom registry file path (overrides user cache and bundled registry)
--baseline	none	Compare scan results against a named baseline (latest selects most recent)
--rules-dir	none	Load additional detection rules from this directory (bundled community rules still apply)
--offline-registry	off	Use bundled registry only, skip user cache
--policy	auto-discover	Path to a governance policy file; auto-discovers .mcp-audit-policy.yml in cwd/repo root when omitted
--verify-hashes	off	Download and verify package hashes against registry (requires network)
--sast	none	Path to MCP server source code to scan with Semgrep SAST rules
--include-extensions	off	Also scan installed IDE extensions for security issues

mcp-audit dashboard flags

Flag	Default	Description
--path	auto-detect	Directory to search for MCP configs
--port	8088	HTTP port for the local dashboard server
--connect	off	Include live-connection findings in the dashboard
--no-open	off	Generate the report without opening a browser tab

GitHub Action

mcp-audit ships as a composite GitHub Action that you can drop into any repository with a single workflow addition. It installs mcp-audit, runs a full scan against your MCP configs, uploads findings to the GitHub Security tab as SARIF, and writes a findings summary to the job summary page. The build fails only when findings at or above your chosen severity threshold exist — making it easy to adopt incrementally (start with severity-threshold: high, tighten to medium once you've cleared existing issues).

Minimal setup

Add this workflow to .github/workflows/mcp-audit.yml in your repo:

name: MCP Security Scan

on:
  push:
    branches: [main]
  pull_request:
    branches: [main]

jobs:
  mcp-audit:
    runs-on: ubuntu-latest
    permissions:
      security-events: write
      contents: read

    steps:
      - uses: actions/checkout@v4

      - name: Run mcp-audit
        uses: adudley78/mcp-audit@main
        with:
          severity-threshold: high
          upload-sarif: 'true'

The permissions: security-events: write block is required for SARIF upload on public repositories. Without it the upload step will fail silently.

Action inputs

Input	Default	Description
severity-threshold	high	Fail the build if findings at or above this level exist (critical, high, medium, low, info)
format	sarif	Output format (sarif, json, terminal)
config-paths	(auto-discover)	Single MCP config file path to scan
baseline	(none)	Baseline name for drift detection
upload-sarif	true	Upload SARIF results to the GitHub Security tab

Action outputs

Output	Description
finding-count	Total number of findings
grade	Letter grade (A–F)
sarif-path	Path to generated SARIF file

More examples

See examples/github-actions/ for:

• basic.yml — visibility-only, never fails the build
• strict.yml — fail on any MEDIUM or higher finding
• with-baseline.yml — drift detection against a committed baseline

Full reference, troubleshooting, and baseline setup instructions: docs/github-action.md.

Pre-Commit Hook

mcp-audit ships as a pre-commit hook, catching MCP misconfigurations before they land in the repository. The hook fires only when a JSON file is staged — no false triggers on Python-only or markdown-only commits — and exits 1 to block the commit when findings at or above your chosen severity threshold exist.

Minimal setup

Add this to your .pre-commit-config.yaml (replace rev with the latest release tag):

repos:
  - repo: https://github.com/adudley78/mcp-audit
    rev: v0.1.0  # Replace with the latest release tag
    hooks:
      - id: mcp-audit

Then install the hooks:

pip install pre-commit
pre-commit install

The hook uses --severity-threshold high by default. To lower the bar to MEDIUM, override args:

hooks:
  - id: mcp-audit
    args: [scan, --severity-threshold, medium]

Note: pass_filenames: false is set intentionally. pre-commit would otherwise pass individual staged JSON filenames to the command, but mcp-audit scan requires full config files discovered through its own client-aware logic. The hook re-scans all MCP configs (not just staged ones) each time it fires.

See examples/pre-commit/ for ready-to-copy config patterns and docs/pre-commit.md for the full reference.

Development

git clone https://github.com/adudley78/mcp-audit.git
cd mcp-audit
uv sync --all-extras

uv run pytest                        # Run all 1,361 tests
uv run ruff check src/ tests/        # Lint
uv run bandit -r src/                # Security audit of the scanner itself

Known limitations

This tool is in early development. See GAPS.md for known detection gaps, untested areas, and planned improvements.

Support

If mcp-audit saves your team time or prevents a security incident, consider sponsoring the project on GitHub.

Sponsorship funds ongoing MCP attack-pattern research, false-positive tuning, new detection rules, and timely releases. Every dollar goes to development time — there is no legal entity, no paid tier, and no gated features. All work remains Apache 2.0 and available to every user.

Not in a position to sponsor? You can help just as much by:

• Opening issues with real-world MCP configs that produce false positives or misses
• Contributing rules — the policy-as-code engine accepts community YAML rules
• Starring the repo so other teams can find it

License

Apache License 2.0 — see LICENSE.