This open-source extension lets you rewrite your X algorithm using a local LLM, and it healed my timeline

Social media has gotten increasingly annoying over the years. With AI in the picture now, it’s even worse. My social feeds are filled with slop I never really asked for. My Instagram is filled with AI fruits having breakups, TikTok is filled with AI voices reading Reddit threads over Subway Surfers, and X is filled with all of it at once. With platforms like Instagram and TikTok, I can easily scroll to the next since the content that largely frustrates me is in video form, which takes a second to identify and a flick to escape. The cost of a bad recommendation is essentially about half a second.

X is a platform I unfortunately can’t escape. Not only do I need to browse through it daily because of the nature of my work, but it’s also the place where I find nerds and geeks like me. With X, ignoring the slop isn’t as easy as on other platforms. After trying everything, I came across an X post (ironic, I know) about an open-source extension that promised to fix exactly this. After going hands-on with it myself, my only regret is not giving it a shot sooner.

Bouncer filters your X feed using plain language

Tell it what you hate, in your own words

If you’ve tried to find a solution to the problem I just described above, chances are you’ve come across a setting called Mute words within X. This feature lets you add words, phrases, usernames, hashtags, and even emojis to a mute list. Every time X comes across a post containing something on your list, it hides it from your timeline and notifications. The feature supports a duration, meaning you can choose to mute targeted notifications for a day, a week, forever — it’s completely up to you!

You can also scope it to notifications, your timeline, or both. For instance, if the word “AI” is really bothering you one day, you can simply add it to your mute list for 24 hours and get some distance from it without committing to never seeing it again. But if there’s one thing I know as a writer, it’s that there are many ways to say the same thing without ever using the word you muted. In the example I just used, it’s extremely simple to talk about AI without ever typing AI.

This is exactly where Bouncer comes in. Instead of a list of words to block, you describe what you don’t want to see in plain English. For instance, AI, crypto, rage politics, engagement bait, etc. Bouncer reads each post as it loads, decides whether it matches, and hides the ones that do. Instead of judging simply on words, Bouncer judges the meaning within.

Once installed, Bouncer embeds itself straight into X’s interface. You’ll see a small panel in the right-hand column, just above the Premium upsell, listing whatever you’ve chosen to filter. From that box you can add filter phrases, toggle AI-slop detection, and hit “View filtered” to see everything it’s pulled. There’s nothing to launch and nothing to switch to. You open X like you always do, and Bouncer is already there!

Under the hood, an AI model does the filtering. You can point Bouncer at a cloud provider (OpenAI, Google Gemini, Anthropic, OpenRouter) or even run a model locally, entirely inside your browser. The latter means no post on your X feed will so much as leave your machine! And instead of hiding your posts away and calling it a day, there’s a View Filtered button you can hit to see every post Bouncer removed, along with the reason it flagged each one. For example, I had AI in my block list, and it hid a post about something a user vibecoded with the following reason:

The post mentions “vibecoding” an app that uses a “Claude agent,” which refers to a popular AI model by Anthropic.

This particular post is an excellent example of why Bouncer works way better than X’s native filtering. This was the post’s caption:

I vibecoded a small app that blinks the capslock led on the MacBook to indicate Claude agent status

Notice how there’s absolutely no mention of AI, LLMs, models, etc. Similarly, another post Bouncer filtered out only mentioned Anthropic in the image! There was no mention in the caption, but Bouncer read the image itself and caught it anyway!

Bouncer can also hide AI slop

It can also sniff out the robots

Beyond the plain-language filters, Bouncer has a dedicated AI-detection mode. It switches on automatically when one of your filter phrases asks for AI content — type something like “AI slop” and it starts flagging AI-generated text and images on its own, no keyword required.

Within Bouncer’s settings, you’ll find three confidence sliders that control how aggressive this detection gets: one for posts, one for comments, and one for images. By default, mine sat at 60% for posts, 50% for comments, and 90% for images.

Now, something I know very well is that AI-text detection is yet to be completely accurate. Yes, it’s gotten a lot better with time, but no detector is airtight, and the false positives don’t disappear.

Bouncer knows this, and it labels catches “Looks like AI text” rather than declaring them AI.

That said, a short post gives it very little to go on. In my week with it, it caught obvious engagement-bait slop reliably but got twitchier on borderline posts. Even imperfect, it’s a handy thing to have running in the background, and the “view filtered” panel means a wrong guess is one click from being undone.

Bouncer earns a permanent spot in my browser

While the extension doesn’t do anything extraordinary, it very quickly rubbed off on me and earned a permanent spot in my browser! It won’t fix X, and in fact, nothing I install will ever change what the platform is. However, for a free, open-source extension, it does a remarkable job of making my corner of X scrollable again.