How Do You Check What's Filling Your Cursor Context?
Look at the context ring next to the prompt input, which shows how full the window is. Click it to open a breakdown tray that splits the tokens used into categories: system prompt, tools, rules, skills, MCP, subagents, summarized conversation and conversation. It tells you what's taking the room before you decide what to cut.
We've been working for a while. Before continuing, summarize in 10 lines or fewer: - The goal and the current state - Decisions we've made and why - Constraints I've given you - What's left to do, with file paths I'll paste this into a new chat, so don't assume anything else will carry over.
- System prompt
- Tools
- Rulesloads every time
- Skillsdescriptions load every time
- MCPcatalogs load every time
- Subagents
- Summarized conversationdetails can be lost
- Conversationyou control this per chat
- 256kStandard window
- 500kLong context limit
- 2xRate above 256kinput billed at twice the standard rate
Try it now
- 1In any active chat, click the context ring next to the prompt input to open the tray.
- 2Hover a segment or row to highlight it, and note how much **Rules, MCP and Skills use compared with Conversation**.
- 3If the fixed categories are large, disable an MCP server you don't use, in Customize, and open a new chat. It worked if the starting usage is lower.
Use it when
When an agent starts repeating itself or forgets instructions, before starting a large task in an existing chat, and whenever you've added new rules or MCP servers. Look at it after your setup changes, because the baseline cost affects every chat afterward.
Skip it when
Don't watch it constantly during short tasks. Usage rising is normal, and Cursor handles compression. Save the attention for long sessions. It also won't explain billing, since usage and prepaid credit are shown elsewhere, under Usage at grok.com/dashboard on The X Subscription.
Why it works
Every chat shares a fixed context window with the model, and files, tool results and messages all fill it. When it's nearly full, Cursor compresses older parts of the conversation into a summary to make room. A summary is shorter by design, so details can be lost. That can show up as an agent that forgets a constraint. The tray separates what you control per chat, like conversation and attached files, from what loads every time, like rules, skill descriptions and MCP catalogs.
Example
Say an agent has started ignoring a limit you gave it an hour ago. You click the ring and see the window is nearly full, with a large Summarized conversation segment. Your constraint may have been compressed away. Rather than repeating it, you send the prompt above, copy the summary into a new chat, and carry on with a clean window.
Going further
To compress on demand, run /summarize. Cursor's CLI docs and release notes describe it, and /compact also triggers it. Subagents are the other big saver, since long searches then happen in their own context. For control, the preCompact hook observes compaction, which lets a team log when it happens. For the model side, the window size depends on your choice: Grok 4.7 has a 256k standard tier with long context up to 500k, billed at 2x above 256k, so a bigger window also costs more. Cached reads are much cheaper than fresh input, so a stable context helps too.
Common mistakes
- Seeing only the percentage and not opening the tray, where the cause is.
- Installing many MCP servers and never checking what their catalogs cost per chat.
- Trusting that a summary kept everything. Important constraints belong in a rule or the next prompt.
- Starting a big task at 90 percent full, so compression happens halfway through it.
Sources
- Prompting agents, Cursor Docs
- Context, Cursor Learn
- Using the CLI, Cursor Docs
- Release notes, Cursor Docs
- Hooks, Cursor Docs
- Available models, Cursor Help
- Models & Pricing, Cursor Docs
- Customize Cursor, Cursor Docs
Checked 6 November 2026. Products change often, so confirm details with the vendor.