Why MCP tool calls burn through tokens
Every MCP-connected tool adds to your context window in two ways: the tool definitions (name, description, input schema) that get loaded so the model knows what's available, and the tool call results that get fed back into the conversation. Connect a dozen servers with verbose schemas and large payloads, and you can burn through a meaningful chunk of your context before the model even answers your question.
1. Trim tool descriptions and schemas
Long, repetitive tool descriptions cost tokens on every single turn, not just once. Keep names and descriptions short and specific, and avoid duplicating information already obvious from the tool name. Flatten input schemas where you can — deeply nested objects cost more tokens to describe than a few flat parameters.
2. Expose fewer, more capable tools
Ten narrow tools ("get_user", "get_user_orders", "get_user_invoices"...) cost more context than two well-designed ones ("get_user_profile" with an optional "include" parameter). Aggregating related actions behind a single tool with parameters usually reduces both the definition overhead and the number of round-trips the model needs.
3. Limit and shape tool output
Return only the fields the model actually needs, not the full raw API response. Paginate large lists instead of dumping everything into one call, and summarize or truncate large text fields server-side rather than relying on the model to do it after the fact — that's tokens you're paying for twice.
4. Scope which tools are even available
Most MCP clients keep every connected server's tools in context for every turn, even if a task only needs one of them. Group related tools by server, connect only what a given workflow needs, and disconnect servers you're not actively using in that session.
5. Cache and reuse results
If a tool call is deterministic or slow-changing (e.g. "list our product catalog"), cache the result at the server layer instead of letting the model re-call it every few turns. Fewer redundant calls means fewer redundant tokens.
How Smia helps
Smia tools run as visual graphs, so it's easy to see exactly what's being sent back to the model at each step — and to add caching, filtering, or aggregation nodes to shrink a tool's output before it ever reaches your assistant's context window.