SlothDesk is a local-first, AI-powered file organizer designed to declutter the Downloads folder by categorizing files and suggesting meaningful filenames, while performing safe, deterministic filesystem operations.

The project aims to simplify the recurring task of organizing digital clutter, giving users control over decisions and data throughout the process.

As a Hacktoberfest Open-Source AI Challenge project, SlothDesk acts as a digital assistant that turns a cluttered folder into an organized download with less repetitive work.

Prize discussions include potential eligibility for Best Use of Gemma, depending on model integration and final implementation results.

The repository hosts the Python implementation, tests, configurations, evaluation code, and documentation under the MIT License, with an openness to feedback and contributions.

Implementation highlights include dry-run previews, a history log, an undo workflow, and an evaluation harness to compare model predictions against ground truth, all aimed at reducing manual file management.

Configurable operation modes via config.json include Ollama (local), Gemini (cloud), OpenRouter (cloud), Apple Foundation Models (on-device), plus a static rules-based mode, each with distinct hardware and privacy implications.

The project traces its lineage to the Kshetrajna project and follows an Observe → Explain → Approve → Act → Recover workflow, introduced as a new application around late October.

Open innovation is central: running open-weight models locally via Ollama gives users control over inference environment, prompts, and policy, while keeping metadata local and ensuring transparency in deterministic steps like categorization and duplicate detection.

Further, local operation avoids sending filenames or snippets to hosted services, supports offline use after installation, and cites optional models and licensing notes for Ollama and Llama 3.1.

A deterministic testing harness with 115 passing tests supports CLI commands, interactive sessions, catalog constraints, and cloud adapters.

A hands-on walkthrough demonstrates real UI interactions where users classify and move files (e.g., client brief to Work, receipts to Receipts), with ambiguous notes potentially left unchanged and user approval required for reclassification.