A $4 rewrite of MCP tool descriptions moved Toolmetry's SQLite success from 34% to 100%

Before the rewrite, the SQLite server's query tool carried this description: "Execute an SQL query against the database. First enumerate available tables using list_tables, then construct your query against those tables." Agents did what it said. They called list_tables before every query, including when the schema was already known from the previous call. Toolmetry counts that as two extra API calls per interaction and doubled latency for no benefit. The description was the only thing requiring that call. Specifically, the word "first".

The replacement string moves the answer into the description: "Execute an SQL query against the database. Available tables: users, orders, products, invoices, and audit_log. Use standard SQL syntax. This tool handles both reads and writes. No need to call list_tables first." The ritual calls stopped. The schema now lives in two places, which is one more than it lived in before. Add a table next quarter and that description is wrong in the way the git server's descriptions were wrong. Those referenced deprecated parameter names, agents passed the old names faithfully, and the calls failed. Toolmetry's stated rule is to update the description in the same commit as the API.

The other SQLite defect was a pair of tools. The read and write query tools were described almost identically, and most failures were the agent calling the read tool when it needed the write one.

The three servers gained at different rates. SQLite had 66.0 points of room to 100 and took all of it. Memory had 38.2 points and took 34.5, or 90 percent. Git had 25.0 points and took 21.7, or 87 percent. After the rewrite the three land within 3.6 points of each other, against a spread of 41.0 points before. Scores that converge that tightly near the ceiling usually mean the task set has been exhausted. The dev.to write-up does not give a trial count per server or name the model driving the agent. I would want both before repeating the 100 percent figure.

For 66 points to transfer, your server has to carry the defects Toolmetry found: two tools a careful human could confuse, a description that implies a prerequisite step, a parameter name the API has dropped. Descriptions that already match the API have nothing to give back. Finding out is the cheap half: list every tool, print each description, read them in pairs, and diff the parameter names against the current API. The write-up prices that at about an hour and under $5. On models it is blunt: "A cheaper model with broken descriptions still fails. An expensive model with broken descriptions still fails."