What MCP tools cost you in context before you call them
Each connected MCP server adds its tool names, descriptions and schemas to context. How that adds up, and how to keep only the servers a project needs.
MCP servers give Claude tools: read a database, open a ticket, query analytics. Each tool comes with a name, a description and a parameter schema, and Claude needs to see those to know the tool exists. That is context spent before anything is called.
Why it adds up
A server with a handful of tools is cheap. A server with dozens of tools, each with a detailed schema, is not, and it is common to have several connected globally, most unrelated to the project at hand.
Keep only what the project uses
- List the servers connected in each project and ask which ones this project really calls.
- Move project-specific servers into that project's configuration.
- Disconnect servers you added once to try.
- Prefer focused servers over sprawling ones when both exist.
Skill or MCP?
Knowledge and procedure belong in a skill, which loads only when used. Live access to an outside system needs MCP. Mixing them up is the most common source of wasted context, as covered in skills vs CLAUDE.md vs MCP. The skill side of the budget is in too many skills installed.
Questions
Do MCP tools use context when I am not using them?
Their definitions do, so Claude knows they exist and how to call them. A server with many tools and long descriptions adds more than one with a few.
How do I reduce MCP context usage?
Connect only the servers a project actually needs, configure servers per project rather than globally, and prefer servers with focused tool sets.
Should something be a skill or an MCP server?
If it needs live access to an outside system, MCP. If it is knowledge or a procedure, a skill, which costs almost nothing until used.