Imbue open‑sources Bouncer

Most people use X with one ranking system: the one X chooses for them. Imbue just published the code for a browser extension called Bouncer that lets a user write their own rules in plain English and hide posts that match them. (github.com) Instead of telling the feed “show me more like this,” Bouncer starts with “don’t show me this.” Its example filters include phrases like “crypto,” “engagement bait,” and “rage politics,” and the extension classifies each incoming post against those topics in real time. (github.com) The way it works is closer to an email spam filter than a mute button. The extension watches the X page for new posts, extracts the text, images, and metadata, sends them to a selected artificial intelligence model, and fades out the posts that match the user’s filter list. (github.com) That matters because X’s built-in controls mostly work at the account or keyword level, while Bouncer is trying to work at the category level. A user can describe a whole class of posts in everyday language, and the model returns both a match decision and a written reason for why it hid the post. (github.com) Imbue did not ship this as one locked-down model. The repository says Bouncer can run with local browser models through Web Language Model, or with cloud models from OpenAI, Google Gemini, Anthropic, and OpenRouter, which means the same extension can be tested with very different moderation setups. (github.com) The local option is the most unusual part. Imbue says some supported models can run entirely inside the browser through Web Graphics Processing Unit acceleration, with model files cached locally, so a user can filter a feed without sending post data to an outside service after setup. (github.com) The codebase is also broader than one Chrome add-on. Imbue’s repository includes a desktop extension for Chrome Manifest Version 3, Firefox, and Safari desktop, plus separate iOS app code, which gives outside developers a ready-made test bed instead of a demo video. (github.com) For researchers, the useful part is not just hiding posts. The repository describes a full pipeline with post detection, batching, caching, provider switching, stored filter phrases, and a “View filtered” review screen, which turns feed ranking into something that can be inspected and changed line by line. (github.com) Imbue’s GitHub page says the company’s broader pitch is that technology should be “loyal to the user,” and Bouncer is that idea applied to a social feed. Instead of asking the platform to rank better, it gives the person reading the feed a second layer of control that sits on top of the platform’s own system. (github.com)