Live data from Hacker News

Show HN: TaskTrace, a local-first activity log that gives LLMs context

tasktrace.com

1–7 of 7 posts

Show HN: TaskTrace, a local-first activity log that gives LLMs context

#1
TaskTrace is a desktop app that continuously records your activity (screenshots, keystrokes, microphone) and turns it into a structured, searchable history of what you’ve been doing.

Everything runs locally on your machine. Local AI models, local database. No cloud, no internet. Your data never leaves your computer.

The data is then summarized and tagged by local LLMs so you get a rich analytics view of how your time is spent. The data is also surfaced by a locally running MCP so that your other agents (Claude, Codex, OpenClaw) can read your activity history and have a better, automatic context of what you've been doing.

Demo: https://youtu.be/GwJy-93vFFI

Show HN: TaskTrace, a local-first activity log that gives LLMs context
tasktrace.com

Re: Show HN: TaskTrace, a local-first activity log that gives LLMs context

#5

Pretty cool idea! I like the "passive" approach. Few questions: Does it automatically take screenshots each X seconds? And which models does it run locally to analyze the images and do the audio transcriptions?

Ty!

It currently takes pictures every 30 seconds and whenever you switch applications.

I use https://huggingface.co/mlx-community/gemma-3-4b-it-qat-4bit to do the chat/image recognition and Qwen/Qwen3-Embedding-0.6B-4bit and Qwen3-Reranker-0.6B-4bit to help in search related features.

For voice I use Apple's SFSpeechRecognizer. I'm thinking of switching that to an OS model, but the memory footprint of the application is already very high.