Skip to main content

Scott Spence's post on MCP Optmization

https://scottspence.com/posts/optimising-mcp-server-context-usage-in-claude-code

Highlights

  • MCP servers all have their own descriptions and prompts for their various functions and tools. However, if you make a lot of these MCP servers and make their descriptions and prompts too detailed, you'll eat your context window before you even get a chance to use it.

  • Architecture fundamentals on the MCP level: if you have multiple MCP servers that do similar tasks, group them under one MCP server. Example, if you have four different MCPs that each handle accessing a specific search engine, group their functionality under one MCP server and use some pattern/approach to choose the search engine that's right for the job.

  • Trim your descriptions. People using your MCP server may need the extended context, but the LLMs don't. Same goes for the descriptions of your parameters in your methods: other developers/people using your MCP server need the full story if they don't deeply understand the domain the MCP handles, but your LLMs don't.

  • Keep an eye on your context usage every so often. Even as you consolidate your MCP servers, think about what additions may cause the token usage to grow again.

Back to top