nachisama/ai-data-extractor

AI Coding Assistant Data Extraction

Extract your own local chat history from AI coding assistants into a single, normalized JSONL format - for fine-tuning, personal analytics, or just backing up years of conversations before an app's local database gets cleared.

Features

Auto-discovers and extracts complete conversation history, including:

• User messages & assistant responses
• Code context (file paths, selections, snippets)
• Code diffs / suggested edits, where the tool records them
• Tool calls and their results
• Timestamps, session IDs, project paths, model names - whatever each

tool's storage actually contains

Supported sources

| # | Tool | Storage | Search location | |---|------|---------|------------------| | 1 | Claude Code | JSONL, one file per session | ~/.claude/projects/**/*.jsonl | | 2 | Codex CLI | JSONL "rollout" files | ~/.codex/sessions/**/rollout-*.jsonl | | 3 | Cursor | SQLite (state.vscdb) | ~/…/Cursor/User/{global,workspace}Storage | | 4 | Windsurf | SQLite, undocumented schema (heuristic) | ~/…/Windsurf/User/{global,workspace}Storage | | 5 | Trae | SQLite + JSONL, undocumented (heuristic) | ~/…/Trae | | 6 | Continue | JSON, one file per session | ~/.continue/sessions/*.json | | 7 | Gemini CLI | JSON, one file per chat | ~/.gemini/tmp//chats/*.json | | 8 | OpenCode | JSON (session/message/part trees) | ~/.local/share/opencode/storage/ | | 9 | Cline / Roo Code (new) | JSON, one folder per task | /User/globalStorage//tasks/ | | 10 | Aider (new) | Markdown transcript per project | /.aider.chat.history.md | Every script searches macOS, Linux, and Windows conventions (~/Library/Application Support, ~/.config, ~/.local/share, %APPDATA%, %LOCALAPPDATA%) automatically - you don't need to tell it which OS you're on.

Why Cline and Aider

They're two of the most-used AI coding tools that the original list left out, and both have a genuinely different (and instructive) storage shape:

• Cline (and its fork Roo Code) is the most popular open-source

autonomous coding agent extension. It stores raw Anthropic-format message arrays per task, so its extractor doubles as the simplest example of parsing that format if you want to add your own tool later.

• Aider is the most popular terminal-only pair-programming tool, and

it's structured completely differently from everything else here: no central database, just a markdown transcript sitting in every project directory. It's included specifically to prove the toolkit generalizes beyond "SQLite or JSONL in one app-data folder."

Quick start

```bash

No dependencies -- standard library only

python --version # 3.9+ required, 3.10+ recommended

Interactive: pick which sources to extract from a numbered menu

python extract.py

Or drive it directly

python extract.py --all python extract.py --sources cursor,claude_code,aider python extract.py --list # just show what's installed, don't extract python extract.py --all --merge # also write all_conversations.jsonl

Shorthand for "extract everything"

./extract_all.sh ```

CLI reference

``` python extract.py [--all] [--sources ids] [--list] [--output-dir DIR] [--search-path PATH ...] [--merge] --all Extract every supported source, no prompts. --sources ids Comma-separated source ids (see table above for names, or run with --list to see them). Skips the menu. --list Report what was found for each source without extracting anything -- a fast, safe preview. --output-dir DIR Where to write JSONL (default: ./extracted_data) --search-path PATH Extra directory to search, on top of the normal OS locations. Repeatable. Mainly useful for:

• Aider, which has no fixed app-data folder and

needs to know where your projects live

• nonstandard install locations for anything else

--merge After extracting, also concatenate everything into all_conversations.jsonl ``` Each extractor can also still be run standalone, same as the original toolkit (python -m extractors.cursor from the project root, or python extractors/cursor.py), which is handy while debugging one source.

Output format

Every run creates timestamped files under extracted_data/: ``` extracted_data/ ├── claude_code_conversations_20260816_143022.jsonl ├── cursor_conversations_20260816_143022.jsonl ├── aider_conversations_20260816_143022.jsonl ├── cline_conversations_20260816_143022.jsonl └── ... one file per source you extracted, plus all_conversations.jsonl if --merge ``` Each line is one JSON conversation: ```json { "messages": [ { "role": "user", "content": "How do I fix this TypeScript error?", "code_context": [ {"file": "/Users/you/project/ai-data-extractor/index.ts", "code": "const x: string = 123;"} ], "timestamp": "2026-01-16T14:30:22Z" }, { "role": "assistant", "content": "The error occurs because you're assigning a number to a string type...", "tool_use": [{"name": "edit_file", "input": {"path": "src/index.ts"}}], "timestamp": "2026-01-16T14:30:25Z" } ], "source": "cursor-composer", "session_id": "c1a2b3...", "project_path": "/Users/you/ai-data-extractor", "name": "TypeScript Type Error Fix", "created_at": 1705414222000 } ``` Fields vary a bit by source (not every tool records code_context, token usage, or project_path) - messages, source, and session_id are the only ones you can always rely on.

How it works

1. Detect the OS and build a list of plausible data roots (Application Support, .config, .local/share, %APPDATA%, etc). 2. Search each root for the tool's known folder name(s). 3. Read the storage - JSONL line-by-line, SQLite via a read-only connection (so a running app never blocks us), or JSON trees, depending on the tool. 4. Normalize whatever's found into the messages[] schema above. 5. Write JSONL, one conversation per line, into extracted_data/. Nothing here ever opens a database for writing, and every reader is wrapped so that one corrupt or locked file can't take down the whole run - you'll get a partial result and move on, not a stack trace.

A note on Cursor, Windsurf, and Trae

None of these publish their storage schema, and it has changed multiple times (Cursor alone has gone through at least three shapes: workspace ItemTable chat, inline composer, and split bubbleId composer). The cursor.py extractor implements all three known shapes explicitly. windsurf.py and trae.py instead use a generic heuristic (extractors/common.py::heuristic_extract_chat_from_kv) that scans chat-related keys and walks the parsed JSON looking for objects shaped like a role + text pair. It's honest best-effort, not a documented format - if a future version changes shape and stops matching, that's expected; adjust KEY_HINTS in the relevant file or send a PR.

A note on Aider

Aider has no central session store - every project directory gets its own .aider.chat.history.md. By default this toolkit scans your home directory plus a handful of common project-root names (projects, code, dev, repos, workspace, src, Documents) up to 5 directories deep, skipping node_modules, .git, and similar. If your projects live somewhere else, point at them directly: ```bash python extract.py --sources aider --search-path ~/client-work --search-path /mnt/data/repos ```

Extending it: adding a new source

Every extractor is a small module with the same two-function interface - copy the simplest one (continue_ext.py is a good template) and fill in: ```python DISPLAY_NAME = "My Tool" SOURCE_ID = "my_tool" def find_installations(extra_paths: list[Path] | None = None) -> list[Path]: """Return the directories/files worth scanning.""" def extract(installations: list[Path]) -> list[dict]: """Return a list of conversation dicts matching the schema above.""" ``` Then register it in extract.py's REGISTRY list. extractors/common.py has the SQLite/JSON/JSONL readers and the two generic heuristics you'll probably want.

Privacy & security

This extracts data from tools running under your own user account. Before sharing or training on it: 1. Scan for secrets: ```bash pip install detect-secrets --break-system-packages detect-secrets scan extracted_data/*.jsonl ``` 2. Review for proprietary code, API keys, and personal file paths - code_context and tool_use fields are the most likely places to find them. 3. Don't commit extracted_data/ to a public repo (it's already in .gitignore). Keep it on encrypted storage if it contains client or proprietary work.

Training use cases

```python from datasets import load_dataset dataset = load_dataset("json", data_files="extracted_data/*.jsonl", split="train") dataset = dataset.filter(lambda x: any(m["role"] == "assistant" for m in x["messages"])) def format_chat(example): return {"text": tokenizer.apply_chat_template(example["messages"], tokenize=False)} dataset = dataset.map(format_chat) ```

Troubleshooting

"No installation found" - the tool either isn't installed, has no chat history yet, or lives somewhere nonstandard. Pass --search-path to point at it directly, or check extractors/.py's SEARCH_DIRS / APP_DIR_NAMES constant and add your path there. Cursor/Windsurf database locked - reads are opened mode=ro specifically so a running editor won't block extraction, but if you still see errors, close the app and re-run. Windsurf/Trae found an installation but 0 conversations - the heuristic key-matching in common.heuristic_extract_chat_from_kv didn't recognize the current storage keys. Run --list to confirm the app dir was found, then inspect state.vscdb's ItemTable/cursorDiskKV keys directly (sqlite3 state.vscdb "SELECT key FROM ItemTable") and add matches to KEY_HINTS.

Disclaimer

This toolkit extracts your own data from AI tools installed on your own machine. You're responsible for:

• having the rights to the extracted data,
• handling any sensitive/proprietary information appropriately,
• complying with each tool's Terms of Service,
• scanning for secrets before sharing or training on the output.

License

MIT - use freely, including for training ML models.