Cursor 3.3's Context Breakdown: Find Out What's Eating Your Tokens

Cursor 3.3 shipped on May 6 with a small feature that solves a big problem. Click the agent's context ring and you now see a breakdown of where every token in your context window is going — split across rules, skills, MCP tools, subagents, file context, and the user prompt. If you have ever watched your agent slow to a crawl on a long task and wondered what was bloating the context, this is the answer.

The reason this matters: most context bloat is invisible. You add a skill three weeks ago, forget about it, and it sits in every session quietly eating 4,000 tokens. The breakdown finally surfaces that cost. Once you see the numbers, the fix is usually one click.

Where to Find the Breakdown

Open the Agents Window and look at the context ring next to the agent's status. The ring fills as your context grows. In Cursor 3.3, that ring is now clickable. Click it once and a popover shows you a stacked breakdown by category.

The categories are: rules (anything in your .cursorrules or rule files), skills (custom skills you have installed), MCPs (each connected MCP server's tool definitions), subagents (the system prompt and tool list of any subagent invocations in this session), file context (files Cursor has read into the session), and user input (your prompts and the agent's replies).

Three Patterns That Show Up Most Often

After diagnosing 30+ sessions with the new breakdown, three patterns emerged. They are the bloat sources you should check first.

Pattern 1: One MCP Eating 30% of Your Context

MCP servers ship with tool definitions, and those definitions get injected into every session that has the MCP enabled. A poorly written MCP server can have 50 tools each with a 200-word description. That's 10,000+ tokens before you have written a single prompt.

If the breakdown shows an MCP taking more than 5% of your context, look at how many of its tools you actually use. Most teams use 3-5 tools from a 50-tool MCP. The fix is to disable the MCP server and re-enable a slimmer fork, or contact the MCP author and ask them to split it.

Click context ring → see "github-mcp: 12,400 tokens (24%)" → 
open Cursor settings → MCP Servers → toggle github-mcp off →
restart agent. Watch context drop by ~12k tokens.

Pattern 2: Stale Skills You Forgot You Installed

Skills are persistent by default. Every skill you installed becomes part of your agent's startup context. If you have been using Cursor since 2025, you probably have skills you do not remember installing. The breakdown will show each one with its token cost.

The fix is simple: open the Skills panel, see the skill names from the breakdown, and uninstall the ones you no longer use. A clean skill list often takes context from 18,000 startup tokens to 4,000.

Pattern 3: Subagents Compounding Without Limits

Subagent calls are the hardest to spot before this feature shipped. Each subagent invocation pulls in the parent agent's relevant context plus the subagent's own system prompt and tool list. Six subagents into a session, you can have 60,000 tokens of subagent overhead in your active window.

The breakdown labels this clearly as "subagents." If it is more than 20% of your context, you probably need to refactor the work. Either consolidate into fewer subagents with broader scopes, or end the session and start fresh now that the parent agent has gathered the relevant findings.

Reading the Numbers

The breakdown shows both absolute token counts and percentage of total. Pay attention to absolute counts when judging individual line items — "35% of context" sounds bad until you realize the total is only 8,000 tokens. Pay attention to percentages when the total is high, since at 180,000 tokens used, every percentage point is real budget.

The user input category is where you want most of the budget going. If your context is 60% rules and skills and 20% user input, your setup is heavy on configuration and light on the actual work. The reverse is healthier.

What This Tool Does Not Do

The breakdown is a snapshot, not a timeline. You see what is in the context right now. You do not see how it got there or what got pruned in earlier turns. If your agent is slow because of how Cursor is summarizing context across long sessions, the breakdown will not catch that — you need the agent debug log for that diagnosis.

It also does not predict cost. Tokens in your context window are tokens you pay for on every turn, but the breakdown shows you what is loaded, not what you have spent across the session. For total spend, check the usage analytics page that shipped on May 4.

A Five-Minute Cleanup You Can Do Now

Open Cursor 3.3, start a new agent session, and send a single prompt. Click the context ring. Look at the categories. Anything that surprises you is a candidate for cleanup. In our testing, the average team running this for the first time cut their startup context by 40%, mostly by removing MCPs and skills they no longer used.