How to spend fewer AI tokens with MCP tools
Published November 6, 2026
Most tokens in an AI that uses tools go to three things: the list of tools it is shown, the raw results tools return, and reading the same data again. Keep the tool list short and searchable, return only the fields that matter, batch independent calls, check for news before reading, and use small models for routine reactions. A good hub does these by default.
Where the tokens go
Every tool an AI app is shown is described in its context on every turn: its name, what it does and its inputs. Fifty tools can cost thousands of tokens before you have asked anything.
Results are the second cost. An API answer for one email can carry headers, ids and formatting that the AI never needs. And scheduled tasks that “check my mail every hour” read the same messages again and again.
Seven habits that keep usage low
These are what Manyplug does for you, without any setting:
- Show only the tools that fit. With many accounts the hub lists the most used tools and lets the AI find the rest with one search tool, instead of describing every tool on every turn.
- Shape results. Empty values are dropped, the AI can ask for only the fields it needs, and an answer over the size budget is cut down instead of flooding the context.
- Batch calls. Independent calls go in one step (manyplug_call_tools), which saves a round trip and the text around it.
- Read synced copies. Messenger chats are synced into the hub, so reading and searching them is fast and does not repeat work on the service.
- Check before reading. A trigger’s “anything new?” check answers {"new": 0} in a few tokens when nothing matched, so a scheduled task does not reread your inbox.
- Follow with cursors. “What’s new” tools take the cursor from the last answer and return only what came after it.
- Use small models for routine work. When the hub wakes your own AI key on an event, it defaults to a small, fast model, caches the unchanging part of the prompt with providers that support it, and stops after a few steps.
What you can do in your prompts
A few habits on your side help as much as anything the hub does:
- Ask for the result, not the process: “Summarize unread mail from clients in five lines” instead of “show me my mail”.
- Name the account when you know it, so the AI does not search all of them.
- For recurring work, schedule a check of a trigger rather than a task that reads everything.
Limits that protect you
Saving tokens is also about not spending them by accident. Triggers that act on their own have a daily limit you choose and pause when they hit it; your own AI keys have their own daily limits; and the hub notices when one chat fires the same trigger over and over.