Bouncer: Open-Source Browser Extension Filtering AI Slop from Twitter in Real Time
Open-source extension Bouncer filters AI slop from Twitter feeds in real time using WebGPU on-device models or cloud LLMs with natural language rules, completel
An open-source browser extension named 'Bouncer' has been officially released to combat the overwhelming surge of low-quality generative AI posts and engagement-bait content flooding Twitter/X feeds in real time. Developed by Millan Philipose (@Millanphilipose) and launched in late September 2026, the tool inspects timeline DOM elements on the fly, evaluating post text and attached media to suppress unwanted synthetic posts before they disrupt the reading experience.
Image credit: @Millanphilipose / GitHub dean2727/bouncer
Real-Time Feed Cleansing via Natural Language Rules and Multimodal Detection
Rather than relying on brittle keyword blacklists or complicated regular expressions, Bouncer allows users to establish filtering boundaries using natural language instructions. Users simply configure conversational prompts such as "crypto", "engagement bait", or "rage politics" directly within the preferences panel. As the timeline scrolls, the extension intercepts incoming DOM elements and classifies whether each post aligns with the specified exclusion criteria.
Unlike conventional blockers limited strictly to textual scraping, Bouncer incorporates multimodal vision evaluation to perform image-aware filtering. This capability allows the extension to detect low-effort AI slop disguised as infographics, reaction memes, or synthetic art. To maintain full transparency, each hidden post features an expandable reasoning panel explaining precisely why it was flagged, accompanied by an adaptive user interface that conforms smoothly to light, dim, and dark display themes.
Flexible AI Backends: From Local WebGPU Execution to Cloud LLMs
Bouncer is engineered to accommodate different privacy requirements and local compute environments through a wide array of AI inference backends.
Its primary technical highlight is complete on-device inference powered by WebLLM and WebGPU. For users running WebGPU-capable browsers, models including Qwen3-4B, Qwen3.5-4B, and Qwen3.5-4B Vision execute entirely within the client's local GPU memory. Because processing occurs inside the browser sandbox, zero post content or user browsing telemetry is ever transmitted to external infrastructure.
For environments prioritizing higher throughput or specialized model architectures, Bouncer provides bring-your-own-key integrations with prominent commercial API providers:
- OpenAI: GPT-5 Nano, gpt-oss-20b
- Google Gemini: Gemini 2.5 Flash Lite, Gemini 2.5 Flash, Gemini 3 Flash Preview
- Anthropic: Claude Haiku 4.5
- OpenRouter: Nemotron Nano 12B VL (free tier), Ministral 3B
- Imbue: Integrated as the platform's default out-of-the-box backend
Addressing model selection in community discussions, Philipose noted that the AI detector relies on a proprietary fine-tune, explaining that in his testing, Jev and other general-purpose models were not effective at AI detection.
Open-Source Availability and Real-World Operational Considerations
Bouncer is distributed under a transparent open-source model hosted on GitHub (dean2727/bouncer), with official browser packages available on both the Chrome Web Store and the Apple App Store. The creator has explicitly pledged that Bouncer is "free to use and always will be."
However, maintaining an active, in-browser classification loop introduces operational tradeoffs that users should weigh:
- Scroll Rendering Overhead: Because DOM elements are intercepted dynamically and processed either via local WebGPU shaders or remote API endpoints, users on constrained client hardware or high-latency connections may encounter brief micro-stutters during rapid continuous scrolling.
- False Positive Risk: Because generative AI detectors operate probabilistically, meticulously crafted human posts or nuanced digital artwork risk occasional misclassification and unintended hiding. Users are advised to calibrate their natural language prompts carefully and inspect the reasoning panel whenever unexpected exclusions occur.