Do you check MCP servers before connecting them to Claude?
retained text
Confession: my process for adding an MCP server is find the repo, copy the npx line into my config, see if it works, move on. I’ve never opened the source of a single one. What got me thinking is that an MCP server can put instructions in its tool descriptions that Claude reads but you never see. Some can also change those descriptions after you’ve approved them. So the thing I connected on Monday might not be the same thing on Friday, and I’d have no clue. Maybe that’s overblown and the popular servers are fine. But I honestly don’t know how worried to be, and I haven’t seen many people talk about what their own habits are. So, what’s your actual routine? Read the code, or at least skim it? Only use servers from known authors or big repos? Pin versions, or let npx grab the latest every time? Run any kind of scanner? Nothing, and it’s been fine? And if anyone’s ever had a server do something weird or sketchy, I’d really like to hear what happened.
document fetched 2026-10-31 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xjbyh1/do_you_check_mcp_servers_before_connecting_them