ChatGPT To Claude MCP

1# chatgpt-to-claude-mcp
2 
3> Turn your ChatGPT export into a personal context profile for Claude and GitHub Copilot.
4 
5Instead 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.
6 
7## How it works
8 
9The pipeline runs entirely locally. It never sends your conversation data anywhere.
10 
111. **Stream** `conversations-*.json` using `ijson` — never loads the full file into memory
122. **Extract** user-only messages (assistant responses are just mirrors of your input)
133. **Score** lines by a weighted blend of word frequency relevance (60%) and Shannon entropy (40%), with context window expansion (±0–3 surrounding lines)
144. **Surface** top proper nouns and bigrams from your history — review and flag noise (old employers, irrelevant contacts, etc.)
155. **Optionally add** domain-specific signal phrases (e.g. `"my game"`, `"my startup"`) to boost relevant context
166. **Re-run** with updated config until the signal looks clean
177. **Read** the dense signal text, then synthesise a profile via the LLM
188. **Install** globally at `~/.config/profile.md`
19 
20Banned phrases persist to `{export_dir}/banned_phrases.json` and signal phrases to `{export_dir}/signal_phrases.json` so they survive reruns.
21 
22## Requirements
23 
24- Python 3.11+
25- [uv](https://github.com/astral-sh/uv) (recommended) or pip
26 
27## Add to Claude Desktop
28 
29`~/Library/Application Support/Claude/claude_desktop_config.json`:
30 
31```json
32{
33  "mcpServers": {
34    "chatgpt-to-claude": {
35      "command": "uvx",
36      "args": ["chatgpt-to-claude-mcp"]
37    }
38  }
39}
40```
41 
42Then restart Claude Desktop and say:
43 
44> "Build my profile from my ChatGPT export at ~/Downloads/chatgpt-export"
45 
46## Add to GitHub Copilot (VS Code)
47 
48`.vscode/mcp.json`:
49 
50```json
51{
52  "servers": {
53    "chatgpt-to-claude": {
54      "type": "stdio",
55      "command": "uvx",
56      "args": ["chatgpt-to-claude-mcp"]
57    }
58  }
59}
60```
61 
62## Install via pip
63 
64```bash
65pip install chatgpt-to-claude-mcp
66chatgpt-to-claude-mcp
67```
68 
69## Tools
70 
71### `process_export(export_dir, banned_phrases?, extra_signal_phrases?)`
72 
73Runs the full pipeline. Returns:
74- `top_proper_nouns` — flag irrelevant ones to ban
75- `top_phrases` — top bigrams, also available for banning
76- `default_signal_phrases` — built-in signal markers
77- `extra_signal_phrases` — your persisted domain-specific additions
78- `signal_lines` / `signal_chars` — size of the dense signal for review
79- `synthesis_prompt` — prompt to use when synthesising
80 
81### `read_signal(export_dir, max_lines?)`
82 
83Read 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.
84 
85### `save_profile(profile_content, export_dir)`
86 
87Writes 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`.
88 
89## Getting your ChatGPT export
90 
91Settings → Data Controls → Export Data. You'll get a zip with `conversations-*.json` files.
92 
93## License
94 
95MIT
96