Cursor Context Window: Check, Clear & Increase [2026]

Cursor uses Claude Sonnet or GPT-4o under the hood, each with different context windows. Sonnet: 200k tokens. GPT-4o: 128k tokens. But what matters is not the size—it is how fast you fill it. Cursor includes MCP tools for file access, codebase search, and execution. These tools return verbose output: a grep with 30 matches can return 10,000-15,000 tokens, even though you only care about 5 matches.

Context window sizes by model

How to check your context usage in Cursor

Open Cursor Composer. Look at the top-right corner—there is a percentage indicator (e.g., "72% context used"). That shows your current consumption. As you add files with @file, run searches with @codebase, or scroll through chat history, the percentage climbs. When it reaches 100%, Cursor starts pruning old messages to make space.

You can also hover over the percentage to see the raw token count: "45,000 / 200,000 tokens used". That tells you exactly how many tokens are left.

Why Cursor context fills so fast

• ✓@file references are verbose — each file includes all content
• ✓@codebase searches return every match plus surrounding context
• ✓Composer history accumulates — every message stays in context
• ✓MCP tool outputs are uncompressed — grep returns 15,000+ tokens for 30 matches
• ✓No deduplication — reading the same file twice counts twice

How to clear the context window manually

When context fills, open a new Composer tab. The new tab has a fresh 0% context window. Click the X on the old tab to close it. Your previous conversation is gone, but you have space to work again. This is a manual workaround—not ideal if you need to reference earlier messages.

How to increase effective context with Token Limits MCP

Token Limits provides eight MCP tools that replace Cursor's defaults: local_read, expand, search, ls, exec, json, diff, map. Each one compresses outputs 60-90% compared to defaults. Install once in your .cursor/mcp.json and every tool call is compressed from that point on.

1. Install: npm install -g token-limits
2. Open ~/.cursor/mcp.json (create if missing)
3. Add: "token-limits": {"command": "token-limits", "args": ["mcp-server"]}
4. Restart Cursor
5. Done. All @file, @codebase, and file operations use compressed tools.

After installation, watch the context indicator. It will climb much more slowly. The same number of files and searches now consume 3-5x fewer tokens.

Real example: Compression in action

You reference 5 files with @file. Default Cursor: ~40k tokens. With Token Limits: ~8k tokens. You search the codebase for "getUserById": default returns 12k tokens (20+ matches with full lines), Token Limits returns 2k. You are at 50k tokens instead of 52k for the same work. Over a 3-hour session, you save 50k-100k+ tokens, extending your runway proportionally.

Comparison: Tools and approaches