Topic Watch: open-source, self-hosted news monitor for productivity

Topic Watch

Topic Watch is a self-hosted news monitor that only notifies you on truly new information using LLM-powered novelty detection.
Fully free and open-source.

Why not RSS readers?

RSS news readers notify you on every single new article on a topic.

This often forces you to re-read the same exact story rehashed 20 times in the periods between truly new information.

Topic Watch tracks a topic and notifies you only when something is genuinely new.

Thus, you save precious time while staying informed on the latest news.

Pipeline

RSS / Bing News / Exa search
        ↓
fetch → dedup → extract content
        ↓
knowledge state + new articles → LLM
        ↓
    has_new_info?
     ├─ no  → record, stay silent
     └─ yes → update state → notify → record

Knowledge state, not keyword matching

Each topic has a knowledge state: a rolling LLM-written summary of what is already established. Every check compares the new articles against that summary to check for truly new details.

Nothing new: the check is recorded, nothing is sent.
Something new: the state absorbs the fact, then a notification goes out with the summary, key facts, and source links.

Cost

You can use any LLM API via LiteLLM - Claude, ChatGPT, Grok, Deepseek, etc. The provider is a single config string.

Roughly $0.0003 per check on a small model (e.g. Claude Haiku 4.5).
Or run Ollama locally and pay nothing.

GitHub link

Fully self-hosted and open-source.
Feel free to share issues and feature ideas!
A star on GitHub is appreciated if you like it :).

Check it out:
github.com/0xzerolight/topic_watch