Find out if the MCP servers you installed are lying to your AI.
Local audit for tool poisoning, rug pulls, tool shadowing, hidden Unicode and leaked secrets, mapped to the OWASP MCP Top 10. Nothing leaves your machine unless you opt in.
Audit MCP servers for tool poisoning, rug pulls and supply-chain risk (OWASP MCP Top 10).
2026-11-10
ดาว
—
เครื่องมือ
—
อัปไทม์
—
🚀 Quick Install
Cursor
Click the button above, or use this deep link: cursor://anysphere.cursor-deeplink/mcp/install?name=mcp-security-guard&config=eyJjb21tYW5kIjoibnB4IiwiYXJncyI6WyIteSIsIm1jcC1zZWN1cml0eS1ndWFyZCJdfQ==
Then run /mcp-audit or ask Claude "are my MCP servers safe?"
Via CLI
claude mcp add mcp-security-guard -- npx -y mcp-security-guard
Claude Desktop
Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):
Your code is covered by other tools. This one checks the MCP servers that inject text into your AI's context.
MCP Security Guard audits every MCP server Claude (and other AI clients) can reach: servers you configured, servers that plugins bring, and servers from Claude Desktop and other clients on your machine.
How It Compares
As of October 2026. Corrections welcome.
Feature
MCP Security Guard
snyk/agent-scan
cisco-ai-defense/mcp-scanner
riseandignite/mcp-shield
Claude Code native plugin
✅
❌
❌
❌
Runtime hooks (pre/post tool-use)
✅
❌
❌
❌
Rug-pull pinning (SHA-256)
✅ Tools, skills, CLAUDE.md
Limited
❌
❌
OWASP MCP Top 10 mapping
✅ Every finding
❌
❌
❌
SARIF output
✅ GitHub Action
❌
❌
❌
CycloneDX BOM export
✅ Agent inventory
❌
❌
❌
Fully local (zero network)
✅ Default
Varies
Varies
Varies
Supply-chain checks
✅ OSV, npm, PyPI
Varies
Varies
Varies
Tool poisoning detection
✅ 26/27 attacks
Varies
Varies
Varies
License
MIT
Apache 2.0
MIT
MIT
Our differentiators:
Native Claude Code integration with a /mcp-audit slash command and runtime protection
Comprehensive rug-pull protection: pins and verifies tool definitions, skills, commands, subagents, and CLAUDE.md files
Every finding mapped to OWASP MCP Top 10 for compliance and remediation guidance
GitHub Action with SARIF for code scanning alerts on pull requests
Agent Bill of Materials (BOM) in CycloneDX 1.6 format for audits and compliance
Privacy-first: everything runs locally by default; network access is explicit opt-in
Use Cases
1. Individual Developers
Protect your local AI setup.
Audit MCP servers before and after installation
Get warned if a server's tool definitions change (rug pulls)
Runtime protection: block credentials from being sent, detect poisoned outputs
Pin trusted servers and skills for continuous verification
# Scan before installing a new server
npx mcp-security-guard scan server-config.json --confirm-launch
# Pin your reviewed servers
npx mcp-security-guard pin
# Check everything is still safe
npx mcp-security-guard audit
2. Pull Request CI Checks
Block risky MCP servers and poisoned skills in code review.
Fail PRs that add:
MCP servers with tool poisoning, unpinned packages, or privileged Docker images
Skills or CLAUDE.md files with instruction overrides or credential exfiltration
Vulnerable dependencies or typosquatted packages
Findings appear in GitHub's code scanning interface with OWASP MCP Top 10 references.
- uses: petrovicistefan/mcp-security-guard@main
with:
fail-on: high
Admins publish one policy (allowed/blocked servers, plugins, pinned versions)
Everyone's plugin enforces it automatically, even offline
Approval flow: developers request, admins approve
Fleet inventory: see what MCP servers your team runs (opt-in with consent)
Compliance evidence: OWASP MCP Top 10 coverage report for audits
What It Checks
MCP Security Guard discovers servers from every place AI clients load them: user, local and project scope, servers shipped inside installed plugins, claude_desktop_config.json, Claude Desktop extensions, organization-managed managed-mcp.json, other clients on the machine (Cursor, VS Code, Windsurf), and claude.ai connectors you've used.
It scans everything a server puts into the AI's context: tools, server instructions, prompts, resources and resource templates.
Check
What It Catches
Tool poisoning
Instruction overrides, "don't tell the user", <IMPORTANT> tags, directives to read secrets (~/.ssh, .env), conversation harvesting, exfiltration via URLs/parameters/Markdown images, HTML comments, encoded payloads
Full-schema poisoning
Same checks on parameter names, descriptions, defaults, enums, required, plus non-schema text in type
Server descriptions referencing another server's tools, tool-name collisions between servers
Rug pulls
Tool definitions or launch commands that changed after you pinned them (SHA-256 per tool). Skills, commands, subagents, CLAUDE.md and hooks of a plugin that changed after pinning (plugins changing files under the same version are flagged high). Re-checked automatically at every session start
Capabilities & score
Per-tool classification (execute, delete, write, egress), unauthenticated remote write access, 0–100 score and A–F grade per server, recommended permission rules
Supply chain
OSV vulnerabilities and malicious versions, typosquats, missing or brand-new packages, install scripts, publisher changes (opt-in network check)
Runtime
Hooks on every MCP call: ask before credentials are sent, warn on injected instructions or credentials in outputs, content-free audit log
Policy
.mcp-security.json approved/blocked servers and hosts, enforced in audits, CI and at session start
Toxic flows
Servers holding all three legs of the "lethal trifecta" (untrusted input, private data, egress capability), on their own or together. A posture note (low or info), not a defect
Agent context
Skills and their scripts, slash commands, subagents, rules, CLAUDE.md and plugin hooks: instruction overrides, "don't tell the user", invisible text, HTML comments addressed to the model, directives to read credential files, commands that upload credentials, download-and-run, encoded execution, infostealer scripts, pre-approved unrestricted Bash
Configuration
Plaintext secrets in env/headers/args/URLs, plain-HTTP remotes, unpinned npx/uvx packages, privileged or unpinned Docker images, pipe-to-shell launches, duplicate names across scopes
Everything runs locally. Nothing is sent anywhere unless you opt in.
OWASP MCP Top 10 Coverage
Every finding is tagged with its OWASP MCP Top 10 id, in reports and in SARIF.
ID
Risk
Coverage
MCP01
Token Mismanagement & Secret Exposure
✅ Plaintext secrets in env, headers, args and URLs. At runtime, asks before a credential is sent to an MCP server and warns when one comes back
✅ 26/27 published techniques detected, including full-schema poisoning, shadowing and name collisions, plus rug-pull pinning
MCP04
Supply Chain Attacks
✅ Unpinned packages, images and git sources; OSV vulnerabilities and malicious versions; typosquats; new packages; install scripts; publisher changes
MCP05
Command Injection & Execution
✅ Flags tools that can execute commands; adversarial_test finds injectable parameters in servers you own
MCP06
Prompt Injection via Contextual Payloads
✅ In tool metadata and, at runtime, in tool outputs (English patterns)
MCP07
Insufficient AuthN/AuthZ
✅ Plain-HTTP remotes; remote servers exposing write or exec tools without authentication; OAuth detection
MCP08
Lack of Audit and Telemetry
✅ Local, content-free audit log of every MCP call, with query_audit_log
MCP09
Shadow MCP Servers
✅ Discovery across user, project, local, plugin and Claude Desktop configs; approved-server policy enforced in audits, CI and at session start
MCP10
Context Injection & Over-Sharing
✅ Conversation and system-prompt harvesting, Markdown-image exfiltration, credentials in outputs, network-egress inventory
What a local tool cannot do (planned for hosted Team plan): org-wide discovery and audit aggregation, OAuth scope review.
Quick Demo (1 Minute)
Try a synthetic offline comparison with no dependencies, account, or API key needed:
git clone https://github.com/petrovicistefan/mcp-security-guard.git
cd mcp-security-guard
node scripts/offline-demo.mjs
The demo checks two saved tool descriptions: one produces instruction-override and sensitive-path findings; the clean counterpart produces no findings. Run the scans yourself and generate an HTML report.
Tools
Tool
Network
Launches
Description
list_mcp_servers
❌
❌
Lists every MCP server configured for this project (user, local and project scope in Claude Code, plus Claude Desktop) with its transport. Read-only; launches nothing.
audit_mcp_config
❌
❌
Static audit of MCP server configuration: plaintext secrets in env/headers/args/URLs, plain-HTTP remote servers, unpinned npx/uvx packages, unpinned or privileged Docker containers, pipe-to-shell launch commands, and duplicate server names. Read-only; launches nothing.
audit_server_tools
❌
✅ Confirm
Connects to selected MCP servers, lists their tools, and checks every name, description and schema for tool poisoning, cross-server tool shadowing, and rug pulls. Never calls the scanned tools. Requires confirm_launch: true.
pin_tools
❌
✅ Confirm
Records a SHA-256 hash of every tool definition in ~/.claude/mcp-security/pins.json for rug-pull detection. Pin only after reviewing an audit. Requires confirm_launch: true.
audit_agent_context
❌
❌
Scans CLAUDE.md, skills, slash commands, subagents, rules and plugin hooks for instruction overrides, requests to hide things from users, invisible text, credential exfiltration, and encoded execution. Read-only; launches nothing.
pin_context
❌
❌
Records SHA-256 of every skill, command, subagent, rule, CLAUDE.md and plugin hook config in ~/.claude/mcp-security/context-pins.json for rug-pull detection in plugins. Pin only after reviewing an audit.
analyze_tool_definitions
❌
❌
Offline analysis of tool definitions you provide (e.g., tools/list output from a server you're developing). Useful in CI or before publishing. Never connects to anything.
check_supply_chain
✅ Confirm
❌
For npx/uvx servers, checks packages on npm/PyPI and OSV database: vulnerabilities, malicious versions, typosquats, install scripts, deprecation, publisher changes. Sends package names and versions only. Requires confirm_network: true.
apply_fixes
Conditional
Conditional
Fixes findings in project files: adds permissions.ask rules (needs confirm_launch), pins unpinned packages (needs confirm_network), or replaces literal secrets with ${VAR} references. Dry run by default; backs up files before writing.
security_dashboard
❌
Conditional
Opens interactive dashboard with scores, grades, findings by severity, OWASP MCP Top 10 breakdown, and pin buttons. scan="config" is read-only; scan="full" requires confirm_launch: true. MCP App hosts render it inline.
generate_policy
❌
❌
Returns a .mcp-security.json policy that approves exactly the currently configured servers. Commit it to flag shadow servers in CI and audits. Read-only; does not write the file.
team_status
✅ Confirm
❌
Syncs organization policy from team backend and shows org name, role, policy version, and fleet reporting status. Requires team API key and confirm_network: true.
request_approval
✅ Confirm
❌
Creates an approval request for a blocked server or plugin when the team policy requires it. Admin decides via CLI. Requires confirm_network: true.
team_report
✅ Confirm
❌
Shows what a fleet report would send (server/plugin names, versions, transports; never paths, args, env, secrets). With confirm_send: true, sends it if admin enabled fleet visibility.
export_bom
❌
❌
Agent bill of materials: inventory of MCP servers, plugins, skills, commands, subagents, CLAUDE.md and hooks with SHA-256, pin status, scores and OWASP evidence. CycloneDX 1.6 JSON or Markdown summary. Static; launches nothing.
query_audit_log
❌
❌
Summarizes the local runtime audit log: MCP tool calls per server/tool, and calls where credentials were sent/returned or outputs contained injected instructions. Log stores hashes and sizes only, never arguments or outputs.
adversarial_test
❌
✅ Confirm
Calls tools with command-injection and path-traversal payloads to find vulnerable parameters. Only for servers you own/operate. Requires i_own_this_server: true and confirm_launch: true. Destructive tools skipped by default.
CI / GitHub Action
Fail pull requests that add risky MCP servers to .mcp.json, and show findings in GitHub code scanning:
name: MCP security
on: [pull_request]
permissions:
contents: read
security-events: write
jobs:
audit:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: petrovicistefan/mcp-security-guard@main
id: mcp
with:
fail-on: high # critical | high | medium | low | info | none
- uses: github/codeql-action/upload-sarif@v3
if: always()
with:
sarif_file: ${{ steps.mcp.outputs.sarif-file }}
To also fail PRs that add poisoned skills, commands or CLAUDE.md, set context: true (second SARIF file in context-sarif-file).
Command line:
# Audit project-level MCP servers
npx mcp-security-guard audit --project-only --format sarif --output mcp.sarif
# For MCP server authors: analyze a saved tools/list result
npx mcp-security-guard analyze-tools tools.json --name my-server
# Scan skills, commands, subagents and CLAUDE.md
npx mcp-security-guard audit-context --project-only --fail-on high
# Pin what you've reviewed
npx mcp-security-guard pin-context
# Scan a server before installing it
npx mcp-security-guard scan server-config.json --confirm-launch
# Test your own server for command injection (calls tools)
npx mcp-security-guard adversarial my-server.mcp.json --server my-server --i-own-this-server --confirm-launch
# Fix what the audit found
npx mcp-security-guard fix --pin-versions # npx pkg → [email protected]
npx mcp-security-guard fix --env-refs --write # secrets → ${VAR} references
npx mcp-security-guard fix --permissions --confirm-launch --write
# Generate a team policy from current configuration
npx mcp-security-guard policy-init && git add .mcp-security.json
Output formats:--format markdown|json|sarif|html (HTML is a self-contained report file).
Exit codes:0 clean, 1 findings at or above --fail-on, 2 usage error.
Interactive Dashboard
security_dashboard is an MCP App that hosts supporting MCP Apps (Claude Desktop, claude.ai, VS Code Copilot) render inline.
Shows every server with its score and grade, findings filterable by severity and server, OWASP MCP Top 10 breakdown, recommended permission rules, and Full scan and Pin buttons.
Web dashboard at /v1/team/dashboard: fleet, approvals, policy, keys, settings
Email alerts (up to 20/hour per org) next to webhook
Compliance evidence: team compliance writes Markdown report with OWASP MCP Top 10 controls, evidence, violations, and pending approvals
Admin actions are CLI only (not MCP tools), so injected text cannot trigger them.
mcp-security-guard team status
mcp-security-guard team report --dry-run
mcp-security-guard team request plugin some-plugin --note "for docs site"
mcp-security-guard team approvals --status pending # admin
mcp-security-guard team approve apr_… --note ok # admin
mcp-security-guard team policy-push policy.json # admin
mcp-security-guard team keys create ana # admin
mcp-security-guard team settings --fleet on --webhook https://hooks.slack.com/… --email [email protected]
Everything above stays free. Free plugin is fully functional. See PRIVACY.md for what is sent.
Benchmark Results
Measured: detects 26/27 attacks from a corpus of publicly documented techniques, with 0 false positives on 13 hard benign samples and on 17 real servers (83 tools).
Techniques covered: tool poisoning and <IMPORTANT> tags (Invariant Labs), line jumping and ANSI escape deception (Trail of Bits), full-schema poisoning (CyberArk), tool shadowing, name collisions, zero-width/Unicode-tag/bidi smuggling, homoglyph names, base64 payloads, HTML comments, exfiltration via URLs/parameters/Markdown images.
npm install
npm run build # bundles to plugin/dist/ (committed, so plugin runs without npm install)
npm test
npm run bench # false-positive gate on real servers (network; Docker images optional)
test/fixtures/poisoned-server.mjs is a deliberately malicious server used by the end-to-end test.
Contributing
Small, reproducible contributions are welcome: benign regression samples, synthetic detection cases, client configuration tests, and clearer examples.
Read-only (except pin files, audit log, and policy-init)
Network only with explicit opt-in
Secrets masked in all output
Static checks reduce risk but don't prove a server safe
License
MIT License
Copyright (c) 2026 Stefan Petrovici
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
Remote servers that require OAuth (most hosted MCP servers) cannot be scanned at the tool level: the scanner cannot reuse Claude Code's tokens. Their configuration is still audited.
About the Author
I'm Stefan Petrovici: passionate about IT, a husband and a father. I built MCP Security Guard on my own. I'm looking for a job.
I build web applications end to end—frontend, backend, APIs and deployment—and I'm happy to work on anything else that needs building. This repository shows how I work: tests, CI, careful documentation and attention to security.
I also build WordPress and WooCommerce plugins, available at pluginsforstores.com.