Designing MCP tool output: raw rows, summaries, or answers with confidence scores
No-Plant-5234 · reddit · 2026-10-01
Building an internal MCP server over databases, the author compares three tool-return designs:
- Raw rows: simple, but explodes the context window on anything bigger than a lookup, forcing the model to summarize rows it never fully saw.
- Server-written summaries: compact, but the server decides what matters and the model repeats the summary as gospel.
- Structured JSON with a confidence score per claim (e.g. 0.97): most useful so far, but the model tends to drop the scores when talking to users.
An open privacy tradeoff: should human-readable tool output include actual data values? Handy for the model, but customer values end up in chat logs. The author asks how others design tool output for data servers and whether anyone got models to keep confidence numbers in replies.
More from coding & agent
- Coding agents waste 10x tokens reading files: 955-job benchmark of a code-graph MCP server — Classeve · 2026-10-01
- Open-Source Runtime Security Layer Keeps Credentials Out of AI Agents — Technical-Spread-368 · 2026-10-01
- Atelico CEO: AI Makes Games Worse—On-Device Small Models Can Fix It — bigdata · 2026-10-01
- Codelab: Build an Apache Iceberg Lakehouse on GCP without external catalogs — jggomezt · 2026-10-01
- Dev shares agent prompt to mine Cloudflare's weekly launches for competitive leverage — threepointone · 2026-10-01
- Google Cloud to release 31 free hands-on episodes on AI agent security this October — dr_cintas · 2026-10-01