Translation MCP team: the hard part wasn't the server, it was getting the model to call it
Pure-Ad7786 · reddit · 2026-09-15
A team shipped a translation MCP server with two tools: smarttranslate runs text through 22 models and returns only the consensus, plus listlanguages. The server was the easy half.
The real problem: the host model already translates, so "translate this into Japanese" gets done inline with a plausible answer and no tool call at all. The same applies to summarization, sentiment, unit conversion—anything the model thinks it already handles. Your tool description is competing with the model's own confidence.
Best fix so far: write the description around what the model can't do (consensus across 22 systems) rather than what the tool does, but it's still inconsistent. The author asks whether anyone has found a description pattern for "this is not the thing you think you can already do."
More from coding & agent
- Dev validates idea on Reddit: dirt-cheap OpenAI-compatible API for open-source models — Extension-Pen-109 · 2026-09-15
- Plus users stretch GPT-6 Astra limits by pairing it with cheap flash models for execution — haider1 · 2026-09-15
- From Ollama to vLLM: a roadmap for scaling LLM deployment — kalyan_kpl · 2026-09-15
- Blender previz plus MiniMax H3 makes AI dance videos far more stable — Hailuo_AI · 2026-09-15
- One group chat, four AI colleagues: a relay workflow for chasing trending topics — lxfater · 2026-09-15
- Stop Building AI Agents in a Line: Why Multi-Agent Workflows Should Be Graphs — nikola_mr64990 · 2026-09-15