---
description: Optimization rules for MCP Server development to ensure fast, responsive tools for LLMs.
globs: "**/mcp-server/**, **/src/mcp/**"
---
# MCP Performance & Optimization
## 1. PAYLOAD MANAGEMENT (Token Saver)
- **TRUNCATION:** Always truncate large text outputs from tools (e.g., > 1000 lines or > 50KB). Provide a clear message: `[Output truncated. Use 'offset' parameter to read more.]`.
- **PAGINATION:** Implement `limit` and `offset` parameters for any tool returning lists or large datasets (files, database rows).
- **FILTERING:** Encourage tools to accept filter parameters (e.g., `glob_pattern`, `search_term`) to reduce data transfer.
## 2. LATENCY & CONCURRENCY
- **ASYNC FIRST:** All I/O operations (file reads, API calls) MUST be asynchronous (`async/await`) to keep the server responsive to multiple tool calls.
- **PARALLELISM:** Use `Promise.all` (JS) or `asyncio.gather` (Python) when a tool needs to fetch data from multiple sources.
## 3. RESOURCE EFFICIENCY
- **CACHING:** Cache expensive resource reads or external API responses (e.g., specific file contents, query results) with a short TTL if frequently accessed.
- **CONNECTION POOLING:** Reuse database or API client connections across tool invocations rather than creating new ones per request.
## 4. NOTIFICATIONS
- **SPARSE UPDATES:** Use resource update notifications sparingly. Only notify when critical data changes to avoid flooding the client and triggering excessive re-reads.
