After a year of experimenting with RAG, web search, llms.txt, and MCP servers, I've found that local documentation works best for coding agents.
The approach: publish docs to npm, let agents explore with `fd`/`rg`.
Why it works: - No remote fetch overhead - Version-matched docs (finally\!) - No N+1 problems like MCP - Agents read only what they need
Benchmarked with gunshi: $1.37 → $0.44 cost, 3:31 → 0:40 time.