Yeda AI Tips · #202

Every Registered MCP Server Costs Tokens on Every Turn

Your unused MCP servers bill you on every single message.

The Model Context Protocol lets an agent talk to external tools — databases, issue trackers, browsers — through a standard interface. It's genuinely useful. But there's a cost that's easy to miss: when a server is registered, its full tool schema (names, descriptions, parameter definitions for every tool it exposes) gets injected into the model's context. That injection happens on every turn, whether or not you actually call the tool. A pile of "just in case" servers is a standing tax on every message you send.

Where the tokens go

A single MCP server can expose a handful of tools, each with a name, a description, and a JSON schema for its arguments. Rendered into context, a typical server adds on the order of a thousand tokens of schema per turn — more for feature-rich servers with many tools. Register eight of them and you're carrying several thousand tokens of tool definitions before you've typed a word. That inflates cost and eats into the context budget you'd rather spend on your actual code.

The insidious part is that idle servers cost the same as busy ones. A server you registered for a one-off task last month is still billing you on today's unrelated conversation.

The mechanism: audit, replace, unregister

Three steps, in order:

Audit what's registered. List your configured servers and be honest about which ones you've actually invoked recently.

   claude mcp list

Replace cloud MCP servers with plain CLIs where you can. An agent that can run shell commands often doesn't need a dedicated server. A gh CLI call reads a GitHub issue without a GitHub MCP server's schema sitting in context; psql or a query script can stand in for a database server. The CLI's help text is loaded only when the agent chooses to run it — not prepended to every turn.

Unregister the rest. If you're not using a server this session, remove it.

   claude mcp remove <name>

MCP server vs. plain CLI

	Registered MCP server	Plain CLI
Schema in context	Every turn, always	Only when invoked
Per-turn token cost	~1k+ tokens each	~0 until used
Setup	Config + running process	Usually already installed
Best for	Rich, stateful integrations you use constantly	One-off or occasional tool calls

MCP isn't the enemy — it shines for integrations you lean on every session. The waste is the long tail of rarely-used servers left registered.

Power moves

Treat eight-plus servers as a red flag. Past roughly that count the schema overhead is a meaningful slice of your context on every message. Prune aggressively.
Audit before a long session. The tax compounds across a long conversation, so clearing dead weight up front pays off most on your heaviest sessions.
Scope servers per project. Register a server in the one repo that needs it rather than globally, so it's not loaded into unrelated work.
Prefer a shell tool over five niche servers. A single command-execution capability plus good CLIs often replaces a stack of narrow MCP servers at a fraction of the per-turn cost.

Resources

Shipping AI or cloud systems? Yeda AI audits and hardens production LLM and agent pipelines. Talk to us · Read the blog