Workloom

Workloom is a local-first desktop activity tracker for knowledge work. It watches the active window, captures screenshots only after filter checks, summarizes them with a local Ollama model, merges them into sessions, and generates searchable history plus daily digests.

Built by Kay @ itskay.co.

What It Does

• Captures activity privately on your machine
• Uses local Ollama models for screenshot understanding and session merging
• Builds daily digests with OpenAI, Anthropic, or fully local Ollama
• Lets you search past work and ask questions about recent activity
• Keeps behavior customizable through config.yaml and user_prompts/

Quick Start

1. Create and activate a virtualenv:

   python3 -m venv .venv
   source .venv/bin/activate
   python -m pip install --upgrade pip
   pip install -r requirements.txt

2. Create a local config:

   cp config.example.yaml config.yaml

3. Pull the recommended local model:

   ollama pull granite3.2-vision

4. Start Ollama:

   ollama serve

5. Fill in config.yaml. For a fully local setup, use digest_provider: "ollama" and point digest_model at a local model or reuse text_model.

6. Start Workloom:

   python run.py

Common Commands

python run.py                 # open the interactive UI
python run.py ui              # re-open / attach the UI
python run.py start           # run headless
python run.py status          # runtime status
python run.py inspect --today # inspect today's captures and sessions
python run.py digest-preview --today
python run.py trends --range week

Inside the UI:

• /inspect
• /search <query>
• /digest
• /trends [week|month|30d]
• /pause
• /resume
• /shutdown

Configuration

There are two main customization surfaces:

• config.yaml Machine-level settings like models, providers, redaction, retention, and chat context.
• user_prompts/ Plain markdown files that control behavior and tone.

Important prompt files:

• digest.md Controls digest structure and section instructions with a simple heading-based format.
• chat.md Controls how the local assistant answers.
• filters.md Controls which apps are skipped or marked low-signal.
• persona.md Controls digest tone and framing.
• summariser.md Controls screenshot interpretation.

If you still have old user_prompts/insight_style.md or user_prompts/digest_format.md files from an older version, they are obsolete and ignored.

Privacy

• Screenshots are only taken after filter checks pass.
• Screenshots are not persisted to disk.
• Capture text is redacted before persistence based on config.yaml.redaction.
• Cloud digest providers only receive redacted session text by default.
• retention_days controls automatic pruning of older captures.

Recommended Defaults

Known-good baseline:

schedule_interval_seconds: 60
vision_model: "granite3.2-vision"
text_model: "granite3.2-vision"
digest_provider: "openai"
digest_model: "gpt-4o-mini"

Alternative local setup to try:

vision_model: "qwen3-vl:4b"
text_model: "qwen3-vl:4b"

Documentation

• Docs Guide
• Implementation Plan

Outputs

• SQLite database: outputs/tracker.db
• Chroma storage: outputs/chroma/
• Log file: outputs/tracker.log
• Runtime state: outputs/runtime_state.json
• Daily digests: outputs/daily_digests/YYYY-MM-DD.md

Notes

• python run.py is the main operator experience.
• inspect.py still exists, but run.py inspect ... is the primary interface.
• The scheduler is designed to keep running even if model or provider calls fail.