chatgpt-to-claude-mcp

Turn your ChatGPT export into a personal context profile for Claude and GitHub Copilot.

Instead of manually writing a system prompt, this tool mines your actual conversation history — your word patterns, recurring topics, and habits of thought — and distils them into a ~3k character profile that meaningfully changes how an AI responds to you.

How it works

The pipeline runs entirely locally. It never sends your conversation data anywhere.

1. Stream conversations-*.json using ijson — never loads the full file into memory
2. Extract user-only messages (assistant responses are just mirrors of your input)
3. Score lines by a weighted blend of word frequency relevance (60%) and Shannon entropy (40%), with context window expansion (±0–3 surrounding lines)
4. Surface top proper nouns and bigrams from your history — review and flag noise (old employers, irrelevant contacts, etc.)
5. Optionally add domain-specific signal phrases (e.g. "my game", "my startup") to boost relevant context
6. Re-run with updated config until the signal looks clean
7. Read the dense signal text, then synthesise a profile via the LLM
8. Install globally at ~/.config/profile.md

Banned phrases persist to {export_dir}/banned_phrases.json and signal phrases to {export_dir}/signal_phrases.json so they survive reruns.

Requirements

• Python 3.11+
• uv (recommended) or pip

Add to Claude Desktop

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

{
  "mcpServers": {
    "chatgpt-to-claude": {
      "command": "uvx",
      "args": ["chatgpt-to-claude-mcp"]
    }
  }
}

Then restart Claude Desktop and say:

"Build my profile from my ChatGPT export at ~/Downloads/chatgpt-export"

Add to GitHub Copilot (VS Code)

.vscode/mcp.json:

{
  "servers": {
    "chatgpt-to-claude": {
      "type": "stdio",
      "command": "uvx",
      "args": ["chatgpt-to-claude-mcp"]
    }
  }
}

Install via pip

pip install chatgpt-to-claude-mcp
chatgpt-to-claude-mcp

Tools

process_export(export_dir, banned_phrases?, extra_signal_phrases?)

Runs the full pipeline. Returns:

• top_proper_nouns — flag irrelevant ones to ban
• top_phrases — top bigrams, also available for banning
• default_signal_phrases — built-in signal markers
• extra_signal_phrases — your persisted domain-specific additions
• signal_lines / signal_chars — size of the dense signal for review
• synthesis_prompt — prompt to use when synthesising

read_signal(export_dir, max_lines?)

Read the dense signal file after approving the config. Signal is scored highest-first — truncating from the bottom loses the least. Returns the text and a frequency summary.

save_profile(profile_content, export_dir)

Writes the synthesised profile to ~/.config/profile.md (with timestamped backup if content has changed) and saves a local copy in {export_dir}/extracted/profile.md.

Getting your ChatGPT export

Settings → Data Controls → Export Data. You'll get a zip with conversations-*.json files.

License

MIT