LLM Dlp — Data skill for Claude Code

llm-dlp

[](https://github.com/yurivenancio30/llm-dlp/actions/workflows/ci.yml) [](https://github.com/yurivenancio30/llm-dlp/releases/latest) [](LICENSE)

**English** · [Português](docs/pt-BR/README.md)

A local proxy that **masks sensitive data between you and the LLM**. The model's API only sees pseudonyms; you, your terminal and the files the agent writes see the real values.

Made for **Claude Code** (stable), with a subscription (Pro, Max) or an API key. The Gemini CLI, Codex, opencode and GitHub Copilot with your own key also work, **in beta** (see [Supported clients](#supported-clients)).

          your machine                                          │ internet
 you ──► Claude Code (or another client) ──► llm-dlp ── masks ──►│──► model API
                                                                │    (sees only pseudonyms)
 you ◄── real values ◄──────────────────── llm-dlp ◄─ unmasks ──│◄── answer with pseudonyms

An example, step by step

You ask the agent: *"write a script that backs up the customer database; the connection is in `.env`"*.

**1. The agent reads `.env`.** This is the file on your disk:

DB_HOST=pg-prd.acme.intra
DB_NAME=crm_clientes
DB_USER=app_crm
DB_PASSWORD=Pr0dS3nha2024
SUPPORT_EMAIL=maria.lopes@acme.com.br

And this is what leaves your machine, the only thing the model's API ever sees:

DB_HOST=hkdzp7s4a.invalid
DB_NAME=db_nr3b5hpi
DB_USER=usr_xyxhy7ug
DB_PASSWORD=SEGREDO-vqxz5dvy
SUPPORT_EMAIL=p.ouhiojea@ds373.invalid

**2. The model answers with the pseudonyms, and llm-dlp swaps them back before the agent acts.** The script saved on your disk, the command that runs and the answer you read have the real values:

pg_dump -h hkdzp7s4a.invalid -U usr_xyxhy7ug -