You plug an MCP server into your assistant, you ask three questions, and the conversation already starts forgetting its beginning. It is not your imagination: most connectors send the AI far more characters than necessary. Here is the exact mechanism, how to measure it on your side, and what can be done about it.
The real culprit: the packaging, not the data
A misleading intuition says the cost comes from the volume of information. In practice, on marketing data, most of the consumption comes from the STRUCTURE that information arrives in.
APIs return tables where every cell is wrapped in its own object, with the field name repeated on every row. On a report of thirty rows and six columns, you are not sending thirty values: you are sending one hundred and eighty small objects, each carrying its column name again, its braces and its quotes.
| What you think you are sending | What actually goes out |
|---|---|
| One row: “/pricing · 1,240 sessions” | One object per cell, with the field name and its punctuation |
| The word “sessions”, once | The word “sessions”, once per row, thirty times over thirty rows |
| A table | A list of objects nested three levels deep |
Measure before optimizing
No need to guess. There is a method that requires no tool and works for any MCP server: ask your assistant to show you the raw response, then count.
Show me the raw response the tool just returned, inside a code block, without rephrasing or summarizing it. Then tell me the number of characters in that response.
Divide the character count by four: you get an order of magnitude of the token count. Repeat the measurement on the three or four tools you call most often, that is where the bill is decided, not on the ones you use once a month.
Four levers, from simplest to most effective
1. Ask for fewer rows
The fastest lever, and the one nobody uses. Most tools accept a limit, a sort and a period. A question phrased with these three details often costs ten times less than an open-ended one.
Give me the 10 pages that lost the most sessions between the last 30 days and the previous 30. Only those 10 rows, sorted by decreasing gap.
2. Turn off the servers you don't use
If your AI client allows it, disable the servers you don't need in the current conversation. Every active server adds its tool list to the context, whether you call it or not. An assistant plugged into eight connectors starts every conversation weighed down.
3. Aggregate server-side, not AI-side
Asking for “the sum of sessions per channel” costs infinitely less than pulling in every row so the AI can add them up itself. When a tool offers an aggregation dimension, use it: you move the computation to where it costs no tokens.
4. Choose a connector that prepares the data
This is the only lever that acts on the packaging itself, and by far the most effective, but it does not depend on you: it depends on how the server was written. A connector can return the API response as is, or reformat it as a table: column names once, then the values. Nothing is lost, nothing is summarized, it is the same data written shorter.
What cannot be optimized from the outside
A question comes up often: can you install an MCP server that would save tokens for the OTHER servers on your account? A “universal saver”, in short.
No, and it is a property of the protocol, not a lack of will. MCP servers are isolated from each other: each one talks directly to the AI client, none sees the others' traffic and none can insert itself in between. A server therefore cannot rewrite a neighbor's responses.
- What works: the server itself preparing its responses, that is Alpative's approach.
- What also works: a proxy that YOU deliberately place in front of your servers, and that replaces their URLs in your configuration.
- What does not work: a third-party server added next to the others, hoping it will optimize them.
Recap
| Lever | Effort | Typical gain |
|---|---|---|
| Limit rows, sort and period | One better-phrased sentence | High |
| Disable unused servers | Two clicks per conversation | Medium, but from the very first message |
| Aggregate server-side | One detail in the question | High on large accounts |
| Connector that prepares the data | The choice of connector | Very high, and permanent |
The first three levers are in your hands today, whatever your connector. The fourth is decided once, at the moment you choose what to plug your assistant into.