You can now see a breakdown of your agent’s context usage!
Click on an agent’s context ring to see how much context is going to rules, skills, MCPs, subagents, and more. This is super useful for spotting when a bucket looks heavier than you’d expect and tightening your setup.
I am a huge fan of this feature 
For some reason it didn’t work for me (clicking the ring in neither editor window nor agent window did anything), until I sent a message in the agent window for the first time (which enables this feature for both the editor window and the agent window). Edge case bug?
Hey @User670839245!
I think the behavior is that you must have sent a chat with v3.3 before you can see this information. After sending a new message, it will work for that chat. Is that what you’re experiencing?
More specifically, I needed to send a chat specifically in the agents window before I can see the context breakdown. I’ve sent plenty of messages in the IDE window.
This is great but it would be further improved if you could drill into each category to see which specific items were contributing to the context usage. For example, context usage by MCP server (and potentially drill into MCP server to see by tool), context usage by rule file/name, context usage by skill.
Without this granularity it is a little difficult to identify particular context that could be reduced. Claude Code /context has provided this granularity for quite some time, I think it would be great to add to Cursor and your version will ultimately be better because it will be visual AND interactive (drill in).
100% this. That was kinda the beauty of the previous interface you could at least see a few of the rules that were being pulled in. If we could click into at least the rules and skills to start that would be awesome
Nice feature. Could we have documentation on what items contribute to each category? Seems no updates about doc site talking about this.
Also noticed a problem: after we do a /summarize, the little circle seems to reflect the current context window (move back a little), hover on it doesn’t show the real percentage, and click on it the break down still shows previous or some stacked-up one. Like 50%, but the circle bar shows like 25%.
Like @Lobob mentioned above. For example: what’s considered “tools”, “MCP”, “skills” and “conversation” when using an MCP that calls some 3rd party API and processes the results?
It could be interesting to be able to drill into “Messages” as well. For example, it could be broken into:
Hey all.
Drilling down into context is something we’re working on and really excited about! More to come.
Thanks for the feedback! Could you please file a bug report and include some screenshots?
Not sure if this will land in the docs, but happy to explain here:
| Bucket | What’s in it | Your control |
|---|---|---|
| System prompt | Cursor’s built-in instructions to the model (how to behave, how to call tools, formatting) | None |
| Tools | The list of tools the model can use and how to call them | Mode (Ask / Plan / Agent etc.) sets the core toolset. A few extras like Web Search and Web Fetch can be toggled in Settings. Core tools (Read, Edit, Grep, Shell, Task…) can’t be turned off |
| Rules | Workspace and user rules that load at the start of the chat. Agent-requestable rules contribute only their description here; the body counts as Conversation if/when the agent reads it | Trim always-on rules; prefer agent-requestable rules with tight descriptions |
| Skills | The catalog of available skills (names + descriptions + paths) loaded at the start. A skill’s body only counts once the agent opens it | Same idea as Rules — fewer/tighter skill entries |
| MCP | Everything your MCP servers inject up front: their instructions and the list of tools/resources they expose. MCP tool results don’t count here, those go to Conversation | Disable MCP servers/tools you don’t need for this chat |
| Subagents | Descriptions of the subagent types you can delegate to | Register fewer subagents |
| Summarized conversation | When the chat gets long, Cursor compresses earlier turns into a summary, which lives here | Automatic or via /summarize |
| Conversation | Everything else: your messages, the agent’s replies and thinking, tool calls and their results, file contents it read, search results, images | Start a new chat, or let it summarize |
Thanks for the detail break down. I saw there’s a simple version of this buried in the docs under “Prompting”.
From what I observe though, “Tools” bucket seems to take account for tool calls or results? It will grow larger if the chat use many tools. If only definition goes there, it should be a constant value?
@Lobob Should only be tool definitions. Just opened up the context breakdown on one of my longer chats, and tool definitions were ~22.1k, same on a new chat.
If you have an example would love to see screenshots!
Yes, different models can have different tool definitions/tools made available! Future iterations of this should let you drill down and see exactly what the difference is.
