context-carry

Import AI conversation exports and serve them via MCP, so any AI application can access your conversation history.

Supported Providers

• ChatGPT — conversations.json from OpenAI data export (DAG linearization)
• Claude.ai — Official Claude data export (conversations.json + projects.json)
• Claude Code — Local ~/.claude/ session data (JSONL)
• Cowork — Claude Cowork sessions with subagent merging

Install

npm install -g context-carry

Or run directly with npx:

npx context-carry import /path/to/export
npx context-carry serve

Usage

Import conversations

# Auto-detect provider from file structure
context-carry import /path/to/chatgpt-export.zip
context-carry import /path/to/claude-export/
context-carry import ~/.claude/

# Force a specific provider
context-carry import /path/to/data --provider chatgpt

# Custom database location (default: ~/.context-carry/conversations.db)
context-carry import /path/to/data --db /path/to/conversations.db

Check status

context-carry status

List conversations

context-carry list
context-carry list --provider chatgpt
context-carry list --search "MQTT"
context-carry list --limit 50 --offset 50

Start MCP server

context-carry serve

MCP Integration

Add context-carry to your AI tool of choice.

Claude Code

claude mcp add --transport stdio --scope user context-carry -- npx context-carry serve

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "context-carry": {
      "command": "npx",
      "args": ["context-carry", "serve"]
    }
  }
}

Gemini CLI

gemini mcp add -s user context-carry npx context-carry serve

Or edit ~/.gemini/settings.json:

{
  "mcpServers": {
    "context-carry": {
      "command": "npx",
      "args": ["context-carry", "serve"]
    }
  }
}

Cursor

Edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "context-carry": {
      "command": "npx",
      "args": ["context-carry", "serve"]
    }
  }
}

VS Code / GitHub Copilot

Edit .vscode/mcp.json in your project root:

{
  "servers": {
    "context-carry": {
      "type": "stdio",
      "command": "npx",
      "args": ["context-carry", "serve"]
    }
  }
}

Note: VS Code uses "servers" (not "mcpServers") and requires "type": "stdio".

Windsurf

Edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "context-carry": {
      "command": "npx",
      "args": ["context-carry", "serve"]
    }
  }
}

Cline

Edit ~/Library/Application Support/Code/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json:

{
  "mcpServers": {
    "context-carry": {
      "command": "npx",
      "args": ["context-carry", "serve"]
    }
  }
}

MCP Tools

| Tool | Description | |------|-------------| | get_user_profile | Synthesized user profile (interests, skills, style) — called automatically on session start | | build_user_profile | Get an overview of all conversations for memory extraction. Supports quick, standard, and deep depth levels | | save_memories | Save extracted memories about the user. Optionally marks conversations as processed | | regenerate_profile | Wipe and rebuild the profile from scratch | | search_conversations | Full-text search across all conversations | | get_conversation | Get a conversation with its full message history | | list_conversations | Paginated list with provider/date/project filters | | list_projects | List all projects/workspaces | | get_project | Project detail with conversation list | | get_stats | Corpus-wide statistics and per-provider breakdown |

How profile building works

When an AI agent calls get_user_profile and no profile exists, it builds one:

1. build_user_profile — Returns all conversation titles and first messages in a single response, grouped by project
2. The AI extracts memories (interests, skills, preferences) from the overview
3. save_memories — Stores the extracted memories and marks conversations as processed
4. get_user_profile — Returns the assembled profile

That's 3 tool calls total for a quick profile, regardless of how many conversations you have.

For richer profiles, use depth="standard" (AI deep-dives ~20-30 interesting conversations) or depth="deep" (deep-dives all conversations).

Development

npm install
npm run dev       # Watch mode
npm test          # Run tests
npm run build     # Production build

Architecture

User's AI export (ZIP/directory)
    ↓
context-carry import <path>     ← CLI (commander)
    ↓
Provider auto-detection → Adapter (ChatGPT/Claude.ai/Claude Code/Cowork)
    ↓
Canonical data model → SQLite + FTS5 (better-sqlite3)
    ↓
context-carry serve             ← MCP server (stdio, @modelcontextprotocol/sdk)
    ↓
Any AI application (Claude Desktop, Claude Code, Cursor, etc.)

License

MIT