Less is More: 4 design patterns for building better MCP servers (www.klavis.ai)

The Model Context Protocol (MCP) is now the de facto standard for agent-tool connectivity across major vendors, but exposing many tools directly to LLM agents creates real production pain: tool hallucinations, decision paralysis, and exploding token costs. Tool definitions (name, purpose, parameter schemas, examples, error handling) can consume 5–7% of a context window before a user prompt — Cursor and Claude Code already expose 18 and 15 tools respectively — so more tools often means worse performance. The “Less is More” principle reframes MCP design: preserve context budget by only delivering what’s relevant when it’s needed. Four practical design patterns address this. 1) Semantic Search: index tool docs with embeddings (example uses SentenceTransformer all‑MiniLM‑L6‑v2 + a vector DB) and retrieve top-k tools per query — simple and scalable but requires tuning. 2) Workflow-Based Design: expose atomic workflow operations (e.g., deploy_project) rather than granular API primitives to cut token use and failure points — ideal for repeatable flows. 3) Code Mode: provide an execute_code sandbox (JSON schema for code+timeout) so LLMs generate programs to run complex, parallelized data tasks (reduces token churn but raises security/debugging concerns). 4) Progressive Discovery: staged discovery (categories → actions → action details → execute) minimizes upfront schema exposure and reduces hallucination for multi-app platforms. Together these patterns enable hybrid MCP architectures that trade a bit of latency for far better reliability, cost-efficiency, and real-world usefulness.