How many tokens do your MCP servers use before you type anything?
Every connected MCP server loads its tool descriptions into the context before the first message. With a handful of servers enabled that adds up fast. Has anyone measured it, and what did you cut: fewer servers, per-project setups, or servers that search for tools instead of loading them all?
Measure before cutting. Start a fresh session with no servers, note the context usage your client reports, then turn servers on one at a time. The difference per server is its fixed cost on every single conversation.
The cost tracks the number of tools and the length of their descriptions, not how useful the server is. A server with one well-described tool is cheap; a server with dozens of tools can cost more than your whole first prompt.
The fixes that work are boring: enable servers per project instead of globally, drop servers whose tools overlap, and prefer clients or servers that look tools up on demand when the catalog is large. Lookup adds a round trip, but it is cheaper than paying for every schema on every turn.
Schemas are only half of it. Long tool results and raw command output fill the context too, so trimming those often saves more than removing one more server.
Listings mentioned
- Fetch MCP Server · agent by modelcontextprotocolA single fetch tool, so its fixed context cost stays small.
- Context Mode Skill · skill by mohitagw15856Filters noisy command output so it does not eat the context you saved.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.