Cursor is now part of your daily work. At this level you stop treating the Agent as one mode and start using its full controls: separating research from execution, extending it with your own tools, letting it remember, and spending the right model on the right task.
Cursor separates thinking from doing. Toggle modes with the picker or Shift + Tab.
Mode
What it does
Plan
Researches the codebase, then writes an editable markdown plan you review before any code is written
Ask
Read-only exploration: answers questions and reads files, makes no edits
Agent
Executes with the full toolset
Plan Mode is the one that changes how complex work feels. It gathers requirements, reads the relevant code, and produces a plan as a markdown document. By default plans save to your home directory; use Save to workspace to commit the plan as a reviewable, shareable artifact before the team writes a line of code.
Model choice is a cost-and-capability decision, and Cursor exposes it directly.
Auto picks a balanced model for the task and runs from a generous included-usage pool, often on Composer, Cursor’s fast in-house model. This is the right default for everyday agentic work.
Max Mode extends the context window to the maximum a model supports and bills at the model’s API token rate, so it burns usage faster. Reach for it only when you genuinely need the larger context or harder reasoning. Depending on your plan, some frontier models and long-context variants require Max Mode.
This is the context window as a budget you manage, not an invisible limit you hit.
The Model Context Protocol (MCP) lets the Agent call external tools: a database, an issue tracker, your internal APIs. Configure servers in:
.cursor/mcp.json (project-scoped, committed for the team)
~/.cursor/mcp.json (your global servers)
A minimal entry runs a local server over stdio and passes it credentials via env:
.cursor/mcp.json
{
"mcpServers": {
"issue-tracker": {
"command": "npx",
"args": ["-y", "your-tracker-mcp-server"],
"env": {
"TRACKER_API_KEY": "your-key"
}
}
}
}
Remote servers use a url field instead of command. MCP supports stdio, SSE, and HTTP transports plus OAuth, and you can install many servers in one click from the in-app Marketplace. Each MCP tool call requires your approval unless you allowlist it.
Memories is Cursor’s auto-memory. Enable it under Settings → Rules → Generate Memories. From then on a background model proposes short, project-scoped facts it observes in your chats (“this repo uses pnpm, not npm”) which you approve before they’re saved. Approved memories are referenced in future sessions and can be reviewed, edited, or deleted from Settings.
Memories are distinct from Rules:
Rules
Author-controlled, committed to the repo, reviewed in pull requests. The team’s shared instructions.
Memories
Model-proposed, you approve them, stored per project by Cursor. Your personal running context.
Cursor’s infrastructure runs on US/AWS, and any frontier-model call sends context to a third-party provider (OpenAI, Anthropic, Google, or xAI). Privacy Mode means neither Cursor nor the model providers train on your code, backed by zero-data-retention agreements, and codebase indexing stores embeddings rather than plaintext. Even so, before connecting M365 or other regulated data via MCP, verify the current DPA, SOC 2, and EU/CH data-residency terms against Cursor’s privacy docs: as of July 2026, Cursor offers no official EU/CH data-residency option. Use .cursorignore / .gitignore to keep sensitive files out of the index.
