Ghost mode is a stringent privacy feature in Cursor that ensures zero data leaves your local machine, operating as a “Local / No-Storage Mode”.

Core Functionality

• When activated, every chat message, code snippet, agent diff, telemetry ping, and trace that would typically go to Cursor’s backend is intercepted locally and discarded.
• Minimal headers, like x-ghost-mode=true, are sent to the server, but no payload containing your code or prompts is persisted or used for model training by Cursor.
• Your code and prompts are sent directly to the Large Language Model (LLM) provider (e.g., OpenAI), subject to that provider’s data retention policies (e.g., typically a 30-day retention window), but Cursor itself does not store this data.

How it Works Under the Hood

• The Cursor client disables internal features that rely on server-side state, such as telemetry, chat storage, and memory/index syncing.
• This strict isolation means that cloud-dependent conveniences are unavailable, including background agents and cloud snapshots, as they require remote server-side state to function.

Why and When to Use Ghost Mode

• Regulated and Ultra-Confidential Codebases: It’s the safest way to use Cursor for projects with sensitive intellectual property (e.g., health, defense), ensuring no off-device storage and potentially meeting “on-prem only” policies.
• Corporate Security Audits: Security teams can use it to verify that Cursor’s traffic only hits model endpoints, as it guarantees local data processing.
• Air-gapped Development: Ghost mode supports fully offline development, especially when combined with local LLMs (like those served by Ollama/Llama.cpp) or your own OpenAI proxy, ensuring no external network calls for data.
• Paranoid Side Projects: If you want absolute control over your data, using Ghost mode means that deleting your local repository erases all traces of your work and chat history within Cursor, as nothing was ever logged remotely.

Feature Trade-offs

• Features that work in Ghost mode: Inline editing, refactoring, AI chat, local Composer and Notepads, locally run Model Context Protocol (MCP) tools, keyboard shortcuts, and CLI operations.
• Features that are disabled or degraded: Background Agents (which require cloud Virtual Machines), memory synchronization, team knowledge sharing, cloud environment snapshots, Bugbot Pull Request (PR) reviews, and cross-device chat history.

Enabling and Verifying Ghost Mode

• You can toggle Ghost mode in Cursor’s settings under Settings → Advanced → Local / Ghost Mode (in older builds, it might be labeled Privacy Mode (legacy no-storage)).
• A restart of Cursor is required for the new sandbox policy to take effect.
• To verify, you can inspect network traffic (e.g., using DevTools) to confirm that every HTTPS request includes the x-ghost-mode: true header and has empty bodies for sensitive data.
• Attempting to use a cloud action (like starting a background agent) should result in a “Feature unavailable in Ghost mode” notification, confirming that the lockdown is active.

Distinction from “Privacy Mode”

• Cursor also has a default “Privacy Mode”. While this mode prevents your code from being used for model training, it still retains some data necessary for features like memories, team sharing, and agent logs on Cursor’s servers.
• Ghost mode is the most restrictive setting, ensuring zero data retention by Cursor and keeping all data on-device.
• This distinction is important because some sources mention that “Privacy Mode” needs to be disabled for background agents. This might refer to Cursor’s default Privacy Mode, which, while more private than not, still involves some data leaving your machine, making it incompatible with the strict “nothing leaves the laptop” policy of Ghost mode. Sources indicate that background agents are available in the broader “Privacy Mode” context where data is not used for training, but still processed remotely. The key takeaway is that Ghost mode represents the highest level of local data isolation.

Best Practices and Gotchas

• Commit Often: Since Ghost mode doesn’t retain snapshots, it’s crucial to commit your work frequently to version control (like Git) to avoid losing chat diffs or progress if your local disk encounters issues.
• Use Rules, Not Memories: Rules, which reside in your repository, are suitable for persistent guidance, as Memories do not sync in Ghost mode.
• Budget Tokens: Without Cursor’s cloud caching, identical prompts will re-query the LLM every time, which can lead to higher API costs for premium models.