MCP server that indexes and searches your past Cursor AI conversations across all workspaces.

Works with any MCP client — Cursor, Claude Desktop, Windsurf, and others.

• Scans ~/.cursor/projects/*/agent-transcripts/ for conversation history
• Builds a persistent full-text search index (SQLite FTS4 via sql.js — zero native dependencies)
• Enriches conversations with metadata (timestamps, mode, branch) from Cursor's internal database
• Exposes 5 MCP tools for searching, browsing, and retrieving past conversations

Add to your MCP client config:

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "cursor-chat-browser": {
      "command": "npx",
      "args": ["-y", "cursor-chat-browser"]
    }
  }
}

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "cursor-chat-browser": {
      "command": "npx",
      "args": ["-y", "cursor-chat-browser"]
    }
  }
}

Then restart your MCP client. The server indexes all conversations on first run (~3s for 1,000 conversations) and uses a persistent cache for fast subsequent startups (~150ms).

Full-text search across all past conversations.

Retrieve full content of a specific conversation.

List all workspaces with conversation counts.

Browse most recent conversations.

Re-scan for new conversations added since the server started.

1. On startup, scans ~/.cursor/projects/*/agent-transcripts/ for JSONL transcript files
2. Parses each file into structured conversations (user messages, assistant responses)
3. Indexes into a persistent SQLite FTS4 database at ~/.cursor/chat-browser/search-index.db
4. Optionally enriches with metadata from Cursor's state.vscdb (timestamps, mode, branch) using the system sqlite3 CLI
5. Serves MCP tools over stdio transport

Only new conversations are parsed on subsequent runs — the index is persistent.

• Node.js >= 20
• Cursor IDE (for conversation data)

Elastic License 2.0