Report

MCP Token Efficiency: Mad Lit vs Notion vs Confluence

28 June 2026 · Model claude-sonnet-4-6 · Measured with Anthropic's count_tokens API

Summary

We measured the token cost of the tool definitions exposed by three MCP servers – Mad Lit, Notion, and Confluence. These definitions are sent to the model on every request, so a leaner schema costs less on every call. For the operations people use most – creating, finding, reading, and editing pages – Mad Lit's tools are consistently the smallest: roughly 10–25× leaner than Notion's and 4.5–6× leaner than Confluence'sfor the same operation.

A leaner schema is only worth it if the model still uses the tools correctly, so we tested that too – and it holds: against Notion, Mad Lit matches tool selection and has a higher search find-rate (see below). This report covers token cost and search-tool quality; it does not assess latency or broader task success.

Method

Token counts come from Anthropic's count_tokens endpoint – the same tokenizer the model uses – run against claude-sonnet-4-6. For each server we captured its complete live tool list and measured the tokens those definitions add to a request.

The per-operation figures below exclude the fixed 497-token “tool-use preamble” the model adds whenever any tool is present (identical for every server), leaving each tool's own schema size. Notion's schema was captured from its hosted MCP (18 tools); Confluence's is a subset of Atlassian's hosted MCP (12 of its 31 tools). Captured June 2026.

Results

Total schema cost

Mad Lit exposes the most tools yet carries the smallest schema. Confluence's 12 page/comment tools alone cost more than Mad Lit's entire 27-tool set. (Confluence here is the Confluence subset of Atlassian's MCP; the full Atlassian server, including Jira, is 31 tools / 13,090 tokens.)

Commonly used operations

Comparing each tool's own definition size for the four everyday operations, in tokens:

On these four, Mad Lit is 10–25× leaner than Notion and 4.5–6× leaner than Confluence. This is a result of the type of inputs: Notion's tools embed its block and property model, and Confluence's create/edit tools document HTML/ADF, panels, and macros, while Mad Lit's accept plain markdown.

What this means in practice– the tool schema rides along on every request, so each new page you create carries its server's entire schema. For the same token cost, that is about 5–6 new pages on Mad Lit for every one on Notion (and ~1.7 for every one on Confluence) – roughly 217 new pages per million tokens of usage on Mad Lit, versus 131 on Confluence and 39 on Notion.

On pay-as-you-go API usage that is a direct cost saving. On Claude or ChatGPT subscription plans, which are not billed per token, it means you reach your usage limit far slower. Creating many pages back-to-back is cheaper on every server, since the schema caches within a short window.

Workflow totals

Fifteen representative tasks were replayed with identical content, counting total tokens across all turns. Averages by task complexity:

Because the tool schema dominates the per-request total, the difference is roughly constant across task complexity.

Does the lean schema still work?

A smaller schema is only an advantage if the model still picks the right tool and uses it well. We tested this for search: the model ran 16 real search tasks against each server's tool schema, with the same retrieval engine held constant, and we measured whether it chose the search tool and whether its query found the right page.

Notion's far larger search schema gave no advantage. The model picked the search tool every time on both, and Notion's find-rate was lower – its queries tended to carry extra words (“pricing notes” instead of “pricing”) that missed, where Mad Lit's leaner description produced tighter queries. The lean schema does the same job just as well, or better.

Find-rate uses the same fixed keyword search for both servers, so it reflects how each schema shapes the model's query, not Notion's own search service. Tool selection and no-match handling are measured directly.

Limitations

• Quality is measured for search only (tool selection and find-rate). Other tools and end-to-end task success are not yet covered, and the report does not measure latency.
• Figures are specific to claude-sonnet-4-6; token counts differ by model.
• Confluence is measured as the Confluence subset of Atlassian's MCP. Tool schemas change over time; these figures reflect each server as of June 2026 and should be re-measured if a tool set changes.