Translate missing i18next strings and measure whether the agent used its tools

Ibrahima D. maintains a JetBrains plugin that gives agents ten MCP tools for translation keys. To test it, they set up a small React app with i18next and 54 English keys, with 92 values missing across French, Spanish and Italian. They ran Claude Code with Sonnet 5.5 headless on a fresh copy each time, with the same prompt that never mentions a tool: three runs with only the JSON files and six with the plugin's tools available. A script then compared every file against the start, checking placeholders, plural pairs and tags, and logged time, cost and turns.

Every run wrote all 92 values with no placeholders or tags lost. Without tools the agent took about 66 seconds and $0.22 a run, and once changed an existing French value it was not asked to touch. In four of six runs with tools it used them, averaging 27 seconds and $0.14. The other two never called a tool. A one-line CLAUDE.md telling it to use the tools fixed that in three more runs.

Claude Code loads MCP tool definitions on demand, so they are cheap in context but easy to ignore. The tool runs also lost existing terminology choices because the agent only saw the missing English text, and wrote keys alphabetically. Twelve sessions on one small project, translations not graded.