I've been adding MCP servers to Claude and Cursor for months — GitHub, a filesystem server, a couple of search servers, a little internal HTTP one I wrote. It works great. Then two things bugged me:

1. Some of those servers have no authentication at all. Anyone who can reach the URL can call my tools.
2. My context window felt full before I even typed a prompt.

Turns out it's not just me. A 2026 analysis of ~7,000 public MCP servers found 41% require no auth, 36.7% are SSRF-vulnerable, and only 8.5% use OAuth. So I wrote a tiny tool to check my own config — and it scored 0 out of 100.

The tool

mcp-audit (https://github.com/alih552/mcp-audit) is a zero-dependency CLI that reads your MCP config (Claude Desktop, Cursor, VS Code, Windsurf, or a plain .mcp.json) and tells you what's wrong. It runs 100% locally — it never connects to your servers or sends your config anywhere.

pipx install git+https://github.com/alih552/mcp-audit
mcp-audit

Here's the kind of thing it flagged on my (deliberately messy) test config:

MCP Audit — ~/.cursor/mcp.json
  7 server(s) - ~13,160 context tokens - score 0/100 (F)

[HIGH] Remote server with no authentication  (internal-api)
[HIGH] Plaintext secret in config (GitHub token)  (github)
[MED]  Unpinned auto-updating executable (npx -y)  (filesystem)
[MED]  Over-broad filesystem root '/Users'  (filesystem)
[LOW]  7 servers ~ 13,160 context tokens loaded every request

What each finding actually means

No auth on a remote server. If your MCP server is reachable over HTTP and doesn't check a token, the model — or anyone who finds the URL — can run your tools. With prompt injection in the wild, the server has to hold the line, not the model.

Plaintext secrets in the config. A GITHUB_TOKEN sitting in .mcp.json leaks through the file itself and through your git history. Move it to an env var or a secret manager.

npx -y / uvx without a pinned version. That silently runs whatever was published most recently. It's a supply-chain risk — pin the version and review updates.

Over-broad filesystem roots. A filesystem server pointed at /Users or $HOME lets the model read and write far more than your project. Scope it to the project directory.

Token bloat. This was the one I didn't expect. Every server loads its tool schemas into every request. Five servers commonly cost 50-75k tokens of context before you type a word — that's real money and real latency. Disable the servers you aren't actively using.

The fix

For the config issues: pin versions, move secrets to env vars, scope filesystem access, and put auth in front of anything remote. There's a full MCP Server Security Checklist here: https://alih552.github.io/mcp-forge/checklist.html

If you're building an MCP server and want it secure from commit one, I also put together MCP Forge Kit (https://alih552.github.io/mcp-forge/) — a secure-by-default starter (bearer + JWT auth, SSRF-safe fetch, rate limiting, validation, tests, CI). But the auditor above is free and MIT, and genuinely useful on its own.

Try it on your setup

pipx install git+https://github.com/alih552/mcp-audit
mcp-audit --json

I'd love feedback on the checks — especially false positives and checks you think are missing. Repo: https://github.com/alih552/mcp-audit