MCP Package Impersonation Auditor

    1

    This skill audits what your MCP servers actually run.

    $24

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    MCP Package Impersonation Auditor

    Example session with this skill installed

    Audit my MCP setup for fake or compromised packages. Config is examples/malicious/claude_desktop_config.json.

    • Read your context and instructions
    • Compiled the mcp package impersonation

    7 servers, 6 critical, 19 high. postmark: npm postmark-mcp is a known-impersonation package (MCP-001); installed copy has a literal Bcc to an outside address (MCP-008), dumps process.env to a request-catcher host (MCP-013), and runs a postinstall that pipes a download into sh. files: @modelcontextprotocol/server-filesytem is one edit from the real package. memory: scope @modelcontextprotocoI uses a capital I. fetcher: mcp-server-fetsh has a .pth hook and an env dump. Also flagged: a sh -c curl|sh launcher, docker --privileged with / mounted, a plain-http remote server, and a lock entry aliasing helper-lib to helper-lib-fork. The skill tells you to stop using the first server, rotate its credentials, and check sent mail for the address.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    In September 2025 a fake postmark-mcp package behaved for fifteen releases, then began copying every email it sent to an outside address. Nothing in anyone's MCP config changed. Scanners that read SKILL.md or tool descriptions would not have noticed.

    This skill audits what your MCP servers actually run. It reads Claude Desktop, Claude Code, Cursor, Windsurf and VS Code MCP configs, works out what each entry launches (npx, uvx, docker, a local script, a remote URL), finds the installed copy in the npx or uv cache, and inspects it.

    What it catches: lookalike and typosquat names (including scope lookalikes such as @modelcontextprotocoI), the right name under the wrong publisher, unscoped twins of official packages, floating versions and npx -y on unpinned names, git and URL installs, lockfile entries with no integrity hash or a resolved URL on plain http, git or another registry, install scripts (and a new release that adds one), Python .pth hooks, a literal address in a bcc or forward field inside mail code, request-catcher and tunnel hosts, whole-environment reads, credential-file reads, child_process and eval patterns, obfuscated code, and the combination of secret access plus network access. Docker entries get checked for --privileged, mounted / or docker.sock, and unpinned tags. With --online (public GETs only) or a saved registry cache it also flags brand-new packages, fresh releases, and a release published by an account that never published the package before.

    scripts/audit_mcp.py is Python 3, standard library only, read-only, and makes no network calls unless you pass --online. It never runs a server and never installs anything. Output has severity, file and line evidence, exit codes for CI, and a plain fix for each rule. examples/ has a malicious impersonator fixture and a clean setup so you can confirm the results yourself.

    This is static pattern checking. A clean result means those signs were absent, not that the code is safe.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.

    Static pattern checks, not a sandbox: a clean result does not prove safety. Packages that are not installed in a cache the script can see are reported as not inspected. Publisher-change detection needs registry metadata (--online or --registry-cache) and is npm only; PyPI gets age checks only. Nested dependencies are scanned only with --deep. Does not read tool descriptions, so it does not catch prompt injection in tool text. The known-bad list is short and tied to public incidents. Fixtures are inert imitations written for testing.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 6 days ago

    Software
    Python 3.9+Postmark MCPSome MCPNode.jsor007 MCP003 MCP006 MCP015 MCP016 MCPresolvesuv (Python tool runner)008 MCP014 MCPcache
    Optional
    Python 3.11+ for uv.lock parsing

    Permissions

    Terminal / Shell
    Read Files
    Network Access

    Allowed Hosts

    registry.npmjs.org
    pypi.org

    Read-only. Reads MCP config files, project lockfiles, and installed package files in the npm and uv caches. No network access unless --online is passed; then only GET requests for public package metadata on registry.npmjs.org and pypi.org. Never runs, installs or modifies anything. Never prints secret values from configs.

    Listed6 days ago

    What's inside

    Frequently Asked Questions

    This skill specifically audits the local code sitting in your `npx` and `uv` caches, identifying typosquats and malicious patterns that standard LLM prompts cannot see because they lack access to your file system and registry metadata.

    The auditor supports configurations for all major MCP-enabled environments, including Claude Desktop, Cursor, VS Code, and Windsurf, by automatically scanning their respective system directories.

    It identifies typosquatting (lookalike names), suspicious publishers, hidden outbound tunnels, unauthorized environment variable access, and known malicious patterns like the postmark-mcp incident.

    No, the skill performs local static analysis and metadata checks; however, using the `--online` flag allows it to verify package age and publisher history against public registries for a more thorough audit.

    Yes, the skill provides specific remediation steps for every threat found, such as how to pin versions, implement lockfiles, or switch to local binaries to mitigate supply-chain risks.

    New in Security