Which AI coding agents and MCP servers run on this laptop or in this repo, and which of them are vulnerable or risky? One file, one command, no account, nothing uploaded but versions and package names.
Download and run
Needs Node.js 22.12 or later. The scanner is a single JavaScript file with no dependencies, served from this site: nothing is installed globally. It is not minified, so you can read it before you run it.
Current version 0.1.0 · 131 KiB · SHA-256 79cc0278cf0edc1f7a246fe25fb3c05c4bc86ba6a33193fdd976ea0fed7db4e8 · version.json
Windows (PowerShell)
Invoke-WebRequest https://threatdeck.net/cli/threatdeck.mjs -OutFile threatdeck.mjs
(Get-FileHash threatdeck.mjs -Algorithm SHA256).Hash # compare with the checksum below
node threatdeck.mjs scan
To update, download the file again; node threatdeck.mjs version prints the version you have. An npm package is planned; until then this download is the way to get the scanner.
Example output
$ node threatdeck.mjs scan
ThreatDeck scan of ~/code/payments-api
Summary: 5 AI tool(s), 4 MCP server(s), 2 high, 2 low, 4 info
Vulnerable or outdated tools
HIGH claude-code 2.1.200 is below the minimum safe version 2.1.251
-> https://threatdeck.net/tools/claude-code/
Risky configuration
HIGH MCP server "search" config contains what looks like a literal credential
Found in env SEARCH_API_KEY. The value is not included in this report or sent anywhere.
at .cursor/mcp.json
LOW MCP server "search" runs npm:search-mcp with no version
npx fetches whatever version is current at every start. Pin an exact version.
Info
INFO Claude Code keeps 212 transcript file(s), 340 MB in plaintext
-> https://threatdeck.net/tools/claude-code/
# Illustrative output: example repo and values; minimum safe version as published on 2026-11-09.
What it checks
AI coding tools & versions
Claude Code, Cursor, Copilot, Codex CLI, Gemini CLI, Windsurf, Cline, Roo Code, Continue, Aider and more, checked against minimum safe versions.
MCP servers in every client
Configs of Claude Code and Desktop, Cursor, VS Code, Windsurf, Gemini CLI, Codex, Roo and Cline, with packages looked up in MCP Server Watch.
Risky configuration
Literal API keys in MCP configs (never printed), unpinned npx servers, bypass-permission settings.
Data at rest
Plaintext credential files and transcript stores, by size only: contents are never opened.
Privacy: exactly what is sent
Never sent: config files, env and header values, arguments, file contents, transcripts. Paths show as ~/….
Sent (plain GET, no cookies or machine id): tool slug + version, MCP package id (registry, name, version), public remote MCP endpoints without query strings, GitHub repo names.
Private hosts, disabled servers and local-path servers are never sent. --offline sends nothing.
To pin a release instead of trusting the latest file, replace the .sha256 download with the checksum above: echo "<sha256> threatdeck.mjs" | sha256sum -c, and update it when you upgrade.
node threatdeck.mjs scan --alert-keys prints follow keys for what it found, for example devtool-claude-code, pkg:npm:@scope/mcp-server. Paste them into a team alert channel and new advisories for those tools and MCP packages land in Slack, Teams or Discord.
Across a team
The fleet view collects the same inventory from every developer machine (opt in with --upload and a fleet token) and shows which tools are below the minimum safe version and which MCP packages have advisories.