About LoosePlume

Fediverse intelligence — built in the open.

LoosePlume listens to public Mastodon conversations, clusters them by topic, and distills each cluster into a short structured report. It reads only what anyone with a browser could read; it never logs in, posts, follows, or interacts.

How it works

Every 30 minutes, LoosePlume polls the trending and hashtag timelines of 17 curated Mastodon instances, plus the reply threads of any status with five or more replies. Posts are embedded with a sentence-transformer and clustered by semantic similarity across instance boundaries. Clusters that clear a wisdom-score threshold — built from reply ratio, boost velocity, thread depth, and cross-instance spread — are fed to a local LLM in a two-stage pipeline (analyst → writer) that produces a title, topic, summary, and support/oppose/mixed key points.

Threads that are just people sharing images or self-promoting (no real disagreement or substantive argument) are detected by the analyst stage and dropped before the writer runs, so the report list stays dense with actual discussion.

What it does not do

• Never posts, boosts, follows, or interacts with any account.
• Never reads private or direct messages — only public timelines.
• Every report links back to the original posts on their home instances.

Source coverage

LoosePlume currently covers Mastodon only. Pleroma, Akkoma, and GoToSocial instances were tested but default to requiring authentication for public timeline reads. Their users' content still reaches LoosePlume when it gets boosted or replied to on covered Mastodon instances via ActivityPub federation.

Instances covered (17)

Hashtags tracked (520)