Designing doc-heavy MCP servers: chunks vs whole files, and keeping tool schemas thin
Southern-Society-713 · reddit · 2026-09-16
A developer building MCP servers for Cursor and Claude Code lays out the core tradeoff for document servers:
- A "read this whole PDF" tool burns tokens and gives the model worse context; semantic search returning a few sections raises quality but requires investment in indexing, chunk boundaries, and citations the agent can trust.
Four concrete questions for people who shipped this:
- What chunk size and overlap actually worked for API specs and design docs?
- Image OCR / video transcripts — same index or separate tools?
- With browser, git, and DB tools installed, how do you stop the MCP tool list itself from eating the context budget?
- Patterns for scoped team vaults (person A shouldn't see person B's uploads by default)?
He has opinions from building in this area but wants to hear what broke in production first.
More from coding & agent
- OpenAI reportedly prepping Codex Replay to run and compare historical task threads in parallel — testingcatalog · 2026-09-16
- LangChain: context engineering is the biggest agent challenge, Deep Agents has it built in — LangChain · 2026-09-16
- Two easy ways to turn an idle Mac mini into a cloud coding agent — iannuttall · 2026-09-16
- HF Jobs OCR script zoo: 28 models, one command, 7 pages for 7 cents — vanstriendaniel · 2026-09-16
- Creator's 3-step system turns 15 manual podcast steps into 8 AI skills, saving 5 hours a week — petergyang · 2026-09-16
- Friends vibe code a personal chat app with retro Windows-style UI, planning a shareable release — floguo · 2026-09-16