I’m trying to understand why the Tools section of my context window is taking up around 21K tokens, even in a brand-new chat where no real work has been done yet. A teammate of mine, in a longer conversation, only has 7.7K tokens in Tools. We’re both on the same project and using the same configuration.
I don’t have any plugins installed, and all my MCPs are deactivated.
In the screenshot below, you can see a brand-new conversation where I only asked it to say “ok,” and the tools were already loaded.
Hey @adangnzlz, thanks for the detailed write-up and screenshot. Here’s some more info on the new context usage breakdown we recently shipped.
The tools bucket contains the serialized definitions (name + description + JSON parameter schema) of Cursor’s built-in agent tools — Read, Write, Edit, Shell, Grep, Task, etc. We send those to the model on every turn, so the agent knows what it can call, which is why they show up before you’ve typed anything.
MCP tools live in a different bucket. MCP server tools are tracked under a separate “MCP” category, not “Tools.” So having no plugins and deactivating MCPs wouldn’t lower the “Tools” number — those two buckets count entirely different things.
Why your teammate’s number is lower a gap that size between two people on the same project is not small, but it could be explained by the following:
Different Cursor version — we roll new and refined tools out gradually, so different builds can have meaningfully different toolsets. Worth comparing Cursor → About Cursor on both machines.
Different agent mode — Agent mode ships the full toolset; Ask is much leaner. If their screenshot was from Ask, that alone could explain most of the gap.
For background, this breakdown UI shipped in Cursor 3.3 and is documented here — it’s surfacing prompt overhead that’s always been there, just newly visible.
If you’re still concerned, you can send a screenshot from your teammates Cursor instances (as well as the respective versions that you’re both on) and we can take a closer look if you’d like.
The “Tools” context bucket contains the definitions for Cursor’s built-in agent tools, such as file reading/editing, terminal commands, search, etc. These are included so the model knows what actions it can take, even before you type a message.
Thanks for clarifying this Kevin.
I think it would be great if we could disable/enable some of those tools.
And as I said above, some additional info within the context summary would have been great.
If you want to make sure the absolute minimum context size is used you can go into the MCP settings and make sure you don’t have any plugins or MCPs, and you can disable the Browser Automation (that alone saves about 700 tokens). But yes there will always be some context that is standard and can’t be removed.
