The problem: your AI agent's package knowledge is months stale I kept hitting the same failure mode while pair-coding with Claude and Cursor: the agent confidently recommends a package, I install it, and only later find out it was deprecated, the repo is archived, or the version it suggested has a known CVE. This is structural, not a model quality issue. An LLM's package knowledge is frozen at its training cutoff — typically 6–18 months stale by the time you use it. In that window a library can get deprecated, hand off maintenance, archive its repo, or pick up a CVE. The model has no way to know any of that happened. A bigger or newer model does not fix this; it just moves the stale cutoff forward a few months. The only real fix is a live lookup at recommendation time . So I built a small thing to do exactly that lookup, and wired it into the agent via MCP so the check happens before the recommendation reaches me.…