kachook/hermes-pulpie

Hermes Pulpie is a plugin that gives the Hermes Agent's web_extract tool a way to clean up web pages before handing them to an AI. When Hermes fetches a page, this tool strips out navigation bars, ads, sidebars, and footers, leaving just the actual content, and it does all of this locally using a small 210 million parameter model rather than calling an external API, so no API key is needed. According to the README, the model matches the quality of a comparable tool called Dripper while being about a third of the size, and it typically cuts the number of tokens needed to represent a page by around 80 percent, which saves cost and context space when the cleaned page is passed to a language model. It automatically detects and uses a GPU (CUDA or Apple's MPS) if one is available, and falls back to the CPU otherwise. Installation is a single pip command that also registers the plugin with Hermes automatically, or it can be installed through Hermes's own plugin installer with a couple of manual dependency steps. Once installed and configured, running a Hermes chat command that extracts a URL will show a footer noting how many characters were saved by the cleanup. Under the hood, the pipeline fetches the page with httpx, runs the Pulpie encoder model to classify each HTML block as content or boilerplate in a single pass, keeps only the content blocks, and converts the result to Markdown with html2text. For pages that rely heavily on JavaScript, like single-page apps or news sites, a separate bundled script uses Playwright to render the page in a headless Chromium browser first, then runs the same cleanup pipeline on the resulting page. The README also documents a full uninstall process covering the different install paths, removing the browser script, freeing the roughly 421 megabyte model cache, and removing the project if it was checked out as a git submodule. The plugin code is released under MIT, the Pulpie library under Apache 2.0, and the model weights under a non-commercial CC BY-NC 4.0 license.