udu-source

Run the udu.py CLI with a thesis and seed companies: `python3 udu.py create "your niche" --seeds a.com,b.com --depth max`. It creates a project, enables the Lead411 contact module, runs the geography, and returns results in the same schema the screening skills read.

Use `python3 udu.py merge --inven inven.json --udu udu.csv --out merged.csv`. It dedupes by normalized domain, lets Inven win ownership and funding conflicts, keeps udu's discovery scores and contacts, and tags each row with its source.

No. The client is pure Python 3 standard library using urllib for HTTP calls, and no MCP server is involved. You only need a udu API key set as UDU_API_KEY or stored in ~/.config/udu/credentials.

udu's founder-owned flag is unreliable on its own and has marked PE roll-ups as founder-owned in testing. The pipeline enforces a mandatory per-company web-check on udu-only rows, and combined mode uses Inven's deal data to override false flags automatically.

Tag at least five clearly on-thesis companies as good and five high-scoring off-thesis companies as bad using `udu.py tag <id> --good ... --bad ...`. This activates udu's AI match score; use only 'bad business model' or 'unrelated' as bad reasons since other reasons do not train the model.

Measured on a live run, udu covers name, domain, and description at 85-100%, headcount around 80%, revenue around 73%, decision-maker name and title around 70%, and personal emails around 45%. Founded year is weaker at roughly 60%, and missing emails fall back to web search plus Verifalia verification.