cursor-db-mcp (MCP Server)

An MCP server for querying, exporting, and analyzing Cursor IDE conversation history from the new workspaceStorage (SQLite) layout.

Highlights

• Query conversations across all workspaces (SQLite state.vscdb under workspaceStorage)

• Search and analyze messages, code blocks, and estimated code changes

• Export data in json, csv, or markdown

• Workspace utilities to list/filter recent workspaces

• Diagnostics tool to discover actual storage keys in your environment

Data sources and fallbacks

• Primary: legacy chat key (if present): workbench.panel.aichat.view.aichat.chatdata

• Heuristics: if the legacy key is missing, the server scans values for structures containing conversations/messages/assistant/role.

• Last resort: synthesizes conversations from aiService.generations (user-side prompts). Optionally, it heuristically pairs items from aiService.prompts as assistant replies to improve completeness.

Note: Depending on your Cursor version, assistant replies may not be locally persisted under known keys. The diagnostics tool can help you discover new keys. If you can share a confirmed key/path containing assistant messages, this server can be adapted quickly.

Installation

Run MCP Inspector for local testing:

The Inspector prints a local URL for interactive testing in your browser.

Configure in Cursor

Basic configuration:

Custom workspaceStorage path:

Available Tools

• list_workspaces

  • Input: { recent_days?: number }

  • Returns workspace list with hash, path, projectPath, lastModified

• get_workspace_conversations

  • Input: { workspace_hash: string }

  • Use list_workspaces to obtain workspace_hash

  • Returns conversations for that workspace

• get_all_conversations

  • Input: { limit?: number }

  • Returns cross-workspace conversations (sorted by updated time)

• search_conversations

  • Input: { query: string, limit?: number }

  • Full-text search in titles and messages

• analyze_conversation

  • Input: { conversation_id: string }

  • Returns stats (message counts, code blocks, estimated code changes)

• export_conversations

  • Input: { format?: 'json'|'csv'|'markdown', conversation_id?: string, conversation_ids?: string[] }

  • Export all or a subset filtered by conversation_id/conversation_ids

• analyze_code_statistics

  • Input: { days?: number, group_by?: 'day'|'week'|'month'|'language'|'workspace' }

  • Aggregated metrics across the selected period

• diagnose_storage

  • Input: { limit?: number }

  • Lists the largest ItemTable keys per workspace to help locate actual chat storage

Ensuring assistant replies in results

• If your environment has a key containing full conversations (including assistant messages), the server will parse it directly.

• If not, the server falls back to aiService.generations (user-only prompts) and tries to pair them with aiService.prompts as assistant replies heuristically. This improves completeness but may not perfectly reflect the original assistant responses.

• To achieve exact assistant outputs, provide the confirmed storage key or path where Cursor persists assistant replies in your version.

Export formats

• JSON: raw conversations array

• CSV: one line per conversation with high-level metadata

• Markdown: per-conversation sections with messages and code blocks

Development

Notes

• SQLite access uses sqlite3. On some systems you may need build tools (e.g., Xcode CLT on macOS) installed.

• Default workspace storage paths:

  • macOS: ~/Library/Application Support/Cursor/User/workspaceStorage

  • Windows: %APPDATA%/Cursor/User/workspaceStorage

  • Linux: ~/.config/Cursor/User/workspaceStorage

Troubleshooting

• No conversations returned:

  • Use diagnose_storage to inspect keys. If no conversation-like key exists, your Cursor version may not store assistant replies locally.

  • The server will still return synthesized conversations from aiService.generations (user prompts). Provide the real chat key to enable full parsing.

• Permission/locking errors:

  • Ensure Cursor is not locking the database when scanning. Try closing heavy operations in Cursor.

License

MIT — see LICENSE.