How-To: Find Which MCP Servers Your AI Agents Launch Fresh From a Registry

After Corelight’s MCP drift report, I went through my own agent configs. Here’s the quick audit.

1. Find the configs

Common spots: Cursor uses ~/.cursor/mcp.json and per-project .cursor/mcp.json. Claude Code uses a project .mcp.json and ~/.claude.json. Other agents have their own; check their docs.

2. Look for launchers that resolve on every start

grep -nE '"(uvx|npx|pipx|bunx)"|@latest' ~/.cursor/mcp.json .cursor/mcp.json .mcp.json ~/.claude.json 2>/dev/null

Any hit means that server can pull new code (or new dependencies) from a public registry next time your agent opens a session, even if you pinned the top-level version.

3. Freeze the ones that hold real credentials

Corelight’s working recipe was: pin to a commit, install once from a hash-pinned lock file, launch the installed binary directly, and turn update checks off. In practice for a Python server:

python3 -m venv ~/mcp/servername
~/mcp/servername/bin/pip install --require-hashes -r requirements.lock

Then point the MCP config’s command at ~/mcp/servername/bin/<server> instead of uvx. See my hash-pinning How-To for generating the lock file.

4. Put re-locking on the calendar

Frozen code ages. Re-lock when the maintainer ships a security fix, review the diff, then reinstall.

God Bless, and Happy hacking!

References & Further Reading

• Your pinned MCP server may not be as pinned as you think (Help Net Security)
• pip: Secure installs (hash-checking mode)
• uv: Locking environments
• Model Context Protocol
• How-To: Pin Your Bots’ Python Dependencies With Hashes (EtherNinja)