When the Simple Solution Beats the Smart One

Spent a few hours this week trying to set up a custom MCP server for ImportYeti inside Claude's Cowork mode. The goal: let Claude query live US import data directly — no copy-pasting, no manual steps.

The MCP server itself worked fine. The problem? Cowork doesn't load custom MCP servers from config files. It uses its own isolated connector system. So no matter what I patched, the tools never showed up.

After going through mcp version conflicts, a JSON parse bug on Windows, a broken bash sandbox, and three different config file locations - I pivoted.

Wrote a 150-line Python script that hits the ImportYeti API directly, dumps results to a JSON file, and lets Claude read it. Took 10 minutes. Works every time.

The lesson I keep relearning: the elegant solution and the right solution aren't always the same thing. Sometimes a script that runs in CMD and writes a file is more useful than a perfectly architected MCP server that never loads.

Takeaway (and you/you're is really just me reminding my future self) If you're building AI workflows, match your tool to your actual constraint, not the constraint you wished you had.

Already used it to pull competitive intel on a competitor - and then run comparative analysis on others using the same or similar manufacturers. It built an interactive live artifact so I can quickly see all the data. That part worked great. KISS - Keep It Simple & Sustainable...

Github for the curious (and the posting process is a nice revisit through everything to reinforce my own learning) https://github.com/FiSimply/importyeti-tools

6 comments