⚠️ 100% AI-Generated Project: This is an experiment in end-to-end AI-driven development, so expect alpha-quality with rough edges. The entire codebase was built by AI coding agents. Zero human-written code. Feedback most welcome!

Gennie‑X extracts and indexes conversations from multiple AI coding assistants into a single, searchable local database. Use it to analyze agent interactions, run semantic search, and generate usage reports. Currently supported sources: GitHub Copilot, Cursor, and Claude Code.

• Multi-Agent Support: Extract conversations from GitHub Copilot, Cursor, and Claude Code
• Web Dashboard: Local web UI to browse workspaces, view analytics, and explore conversations
• CLI Interface: Equivalent command-line tools for automation and scripting
• Semantic Search: Search through your conversation history with AI-powered semantic search (Sentence Transformers embeddings + SQLite FTS5 keyword search)
• Code Metrics Extraction: Extracts code metrics (via lizard and related tooling) to summarize complexity and size across extracted code artifacts
• Plugins: extensible to support new AI coding agents

• Python 3.11 or higher
• uv - Fast Python package manager
• One or more supported AI coding agents installed (GitHub Copilot / Cursor / Claude Code)

1. Clone the repository

   git clone https://github.com/hosamsh/gennie-x.git
   cd gennie-x

2. Install dependencies

   # CPU-only (default, works on all machines)
   uv sync

Important: Gennie-X does not create workspaces. Instead, it discovers existing workspaces from your AI coding agents (VS Code/Cursor/Claude Code). The tool reads the local storage where these agents save their conversation history.

To have workspaces available:

1. Use one of the supported agents (GitHub Copilot in VS Code, Cursor, or Claude Code)
2. Open a project/folder in the agent
3. Start at least one chat conversation with the AI assistant
4. The agent will save this data to its local storage (auto-detected by Gennie-X)

Create the config file by copying or renaming the example config. The defaults work for most setups:

cp config/config.example.yaml config/config.yaml

Launch the web dashboard to explore your conversations visually:

uv run python run_web.py
# Opens at http://127.0.0.1:8000

The web interface lets you browse workspaces, extract chat sessions, view analytics, and explore conversations interactively without needing to use the CLI.

If the web UI shows no workspaces, it usually means the app is pointing at the wrong storage location or there is no local agent data yet. Check the storage paths in config/config.yaml (section extract.<agent>.workspace_storage). If you have not used the agent locally, open any workspace in the agent and start a chat once so it writes local storage, then refresh.

# Get help on all available commands
uv run python run_cli.py --help

# List all available workspaces
uv run python run_cli.py --list

# Extract specific workspace (requires workspace ID and output directory)
uv run python run_cli.py --extract <workspace-id> --run-dir data/<dir-name>

# Extract all workspaces for a specific agent
uv run python run_cli.py --extract --all --agent <copilot|cursor|claude_code> --run-dir data/<dir-name>

# Refresh an existing workspace (incremental sync; only new/changed turns)
uv run python run_cli.py --extract <workspace-id> --run-dir data/<dir-name>

# Refresh an existing workspace (full re-ingest; reprocess everything)
uv run python run_cli.py --extract <workspace-id> --run-dir data/<dir-name> --force-refresh

# Full reset (erase DB, then re-ingest)
rm -rf data/<dir-name>
uv run python run_cli.py --extract <workspace-id> --run-dir data/<dir-name>

# Search through extracted conversations
uv run python run_cli.py --search "your search query" --run-dir data/<dir-name>

If --list shows 0 workspaces, it usually means the app is pointing at the wrong storage location or there is no local agent data yet. Check the storage paths in config/config.yaml (see section extract.<agent>.workspace_storage in the config). If you have not used the agent locally, open any workspace in the agent and start a chat once so it writes local storage, then re-run --list.

• Location: config/config.yaml (loaded by default)
• Example: config/config.example.yaml (copy/compare for defaults)

Sections:

• extract: text shrinking + agent storage/output paths
• web: web UI run directory and port
• search: search mode, thresholds, embedding model + batch size, auto-embed
• token_estimation: token estimation heuristics
• loc_counting: rules for couting lines of code
• logging: log level

You can also edit the config via the web UI.

If you want faster embedding generation, you can enable GPU support:

1. Install a CUDA‑enabled PyTorch build (see requirements-gpu.txt for guidance).
2. Then install the rest of the dependencies.

To choose the right CUDA wheel: (1) check your NVIDIA driver/CUDA capability (e.g., from nvidia-smi), (2) pick the matching CUDA version on the PyTorch install page (e.g., cu118 or cu121), and (3) use that index URL in requirements-gpu.txt.

1. Create src/extract_plugins/your_agent/
2. Implement agent.py with AgentExtractor interface
3. Add metadata.json for display info
4. Add config section to config/config.yaml
5. Test extraction and enrichment

See Agent Extractor Interface Docs

see tests/readme.md

Note: Do not use python -m pytest or activate a .venv manually - use uv run pytest instead.

• Primarily tested on Windows
• Duplicate workspace entries when edited by Claude Code + other vs-code based agents
• Claude Code session names are GUIDs
• Left-side navigation resize briefly on page loads

TBD.

You’re free to use, copy, modify, and redistribute this project.

Gennie‑X reads agent‑specific local storage formats to reconstruct conversations. Those formats and field locations can change between agent releases, and there is no universal standard—so extraction is best‑effort and may require occasional updates to extractors. If you notice missing or misaligned data, please open an issue; contributions to improve extractor robustness are welcome.