Cursor loads MCP tools on demand: what it means for Sume

Cursor says dynamic MCP context cut total tokens 46.9% in sessions that call an MCP tool. For Sume, keep the tool list short and use tools_list.

4 min readSume
All posts

Cursor's September 23, 2026 post says it moved MCP tool descriptions into dynamic context, loaded only when needed, and that this cut total tokens by 46.9% across sessions that called an MCP tool. If you connect Cursor to Sume's hosted MCP, the server's own two-step discovery works with that design: names first, one schema on demand.

The numbers are Cursor's, from its token efficiency post, and they describe Cursor's sessions. This post does not claim a figure for Sume.

What Cursor changed

Cursor's token-efficiency changes (read 2026-10-04)
ItemWhat the post says
MCP toolsMoved into dynamic context, loaded only when needed
Reduction46.9% fewer total tokens in sessions that called an MCP tool
High-frequency toolsStay in static context
Overall7% cost reduction; system prompt trimmed about 66%

How does this help with a big tool list?

Sume exposes assets, generation, crawl, avatar and timeline tools. The hosted MCP docs list discovery tools (mcp_health, tools_list, tools_schema, account_me, catalog_list), and a client can ask tools_list for names and tools_schema for one tool at a time.

With on-demand loading in the client, a session that only needs a crawl read never pays to describe video generation. If your session needs a single tool every turn, that is the case Cursor keeps static, so use the editor's own settings to see how it classifies it.

Tips for a Cursor setup

  • Connect with OAuth and accept mcp:read first; add mcp:write only for paid work.
  • Ask the agent to call tools_list and report the names before it picks one.
  • Do not copy Cursor's percentage into your own reports.
  • Re-read the Cursor post; its defaults may change.

Sources

Related posts

More in Integrations

All Integrations posts

Written by Sume