This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine
TL;DR: I keep seeing the same posts on most of the DEV home feed tabs:
Discover,FollowingandCurated.Wanted to fix it by implementing a simple Mark as Seen browser extension. Then one thing led to another, and I ended up building three projects: one already abandoned, and two ongoing (Feedgrab & Ghostify).
Note for users: This is still under development, so if you see errors, please look for updates: https://github.com/SweaDev/ghostify/releases
Note for judges: v1.0.0 was completed within the submission period; later releases are updates.
Backstory
These days, I spend too much time on DEV.
- Deleted my old X/Twitter account because it gives me a headache! Created a new one only as a placeholder.
- Deactivated my Facebook account, because it's… well, that's a story for another day!
So most of the time I spend online is here on DEV and YouTube.
But the problem with DEV (and YouTube too) is that I keep seeing the same posts in the feed!
Open DEV in the morning: there's that post. Open it at lunch: there it is again. Open it at night: guess who's back! I've read it. I've reacted to it. I may have even commented on it. Still there, staring at me like it hasn't been introduced yet.
So on October 26, I did what any sane developer would do: I opened my editor instead of closing the tab.
Attempt #1: myDEV (RIP 🪦)
The plan was simple:
- Put a ✓ button on every post card, right next to Bookmark.
- Click it, and the post disappears from the feed and goes into a new myDEV tab.
- Done. Go back to reading.
And it worked! For about an hour. Then I opened YouTube and saw the same video for the 14th time this week, and realized myDEV only knew how to read one site, with selectors I hard-coded by hand. The moment DEV changes a class name, it's dead. And for YouTube, I'd have to start over.
It took me a few days to admit it, but today I finally archived it. Rest in peace, myDEV. You were too specific for this world.
Attempt #2: Feedgrab
What I actually needed wasn't a DEV plugin. I needed something that could look at any page and say: "this is a feed, these are the items, and this one is the same item you saw before, even after infinite scroll re-rendered it."
That became Feedgrab, a TypeScript library that:
- detects feeds on a page (structural heuristics, optionally refined by an LLM), or takes CSS/XPath selectors you type;
- lets you confirm or fix its guesses in a Shadow DOM overlay;
- turns the feed into one de-duplicated stream of items that keeps up with infinite scroll;
- never removes anything unless you ask it to.
It runs in browser extensions, headless browsers and on the server, with the same API.
Attempt #3: Ghostify
And then, finally, the thing I wanted in the first place. On top of Feedgrab. With a ghost. 👻
What I Built
Ghostify is a browser extension (Chrome + Firefox) for feeds that keep showing you the same posts.
Hover a post, click the ghost, and it's gone from the feed. For good. Reload, scroll, come back tomorrow: it stays gone.
But it's not deleted. It becomes a ghost. Ghostify keeps a copy, so you can find it again, read it later, or snooze it until the weekend.
That's my take on Ghost in the Machine: the posts you've already seen don't die, they just stop haunting your feed and move into the machine. Quietly. Politely. Until you call them back.
What it does:
- The ghost button shows up on the post under your pointer. Click it, the post vanishes, and an Undo toast gives you 5 seconds of regret.
- The Ghostify page keeps everything in three tabs: Ghosted, Later and Snoozed.
- Snooze a post until later today, tomorrow, the weekend, next week, or any date. When time's up, it shows up in Back from snooze.
- Search your ghosts by keyword, or switch on Semantic and just describe what you remember ("that post about JavaScript adding objects and arrays").
- Auto-ghostify rules in plain English, per site: "recurring weekly discussion threads I have already read". A local decision model checks new feed items against them and ghosts the matches. If it gets one wrong, Not this brings the post back, and that rule never touches it again.
- Add any site. On any page with a feed, click Add this site. Feedgrab proposes the feeds, you confirm them in its overlay, and choose whether they're for the logged-in or logged-out view.
- New tab flavor, if you want your ghosts on every new tab.
Demo
No store listing (yet!), so the demo is: install it, open DEV, start ghosting. It takes about a minute:
- Download a zip from Releases:
ghostify-<version>-chrome.ziporghostify-<version>-firefox.zip. -
Chrome: unzip, open
chrome://extensions, turn on Developer mode, click Load unpacked. Firefox: openabout:debugging#/runtime/this-firefox, click Load Temporary Add-on… and pick the zip. - Open https://dev.to/ logged out, hover a post, click the ghost.
Here's a quick tour. (The screenshots represent what a user would see.)
The toolbar popup lists your recent ghosts. Click the 👁 to unghost one, and it shows up in the feed again:
Click a ghost to see its details, the AI summary and tags:
Write auto-ghostify rules in plain English, per site:
And pick a model for each job:
Code
Ghostify
Ghostify is a browser extension (Chrome and Firefox) for feeds that keep showing you the same posts. Hover a post, click the ghost, and it disappears from the site's feed for good. Ghostify keeps a copy so you can find it again later.
v1.0.0. Ghostify works on the dev.to home feeds out of the box, and you can add any site. It has a full Ghostify page with snooze and search, and optional AI that runs locally with Ollama or with your own OpenRouter, Gemini or OpenAI key.
Features
-
The dev.to logged-out home feeds (Relevant
/, Latest/latest, Top/top/*) are recognized automatically, through feedgrab. -
The ghost button shows on the post under your pointer. Clicking it hides the post and saves it, and an Undo toast lets you take it back.
-
Ghosted posts stay hidden when you reload, and also when they come back…
Feedgrab
A TypeScript library that finds and describes content feeds in web pages (post timelines search results, product grids, article lists) and turns them into one de-duplicated stream of items with exclusion reports (ads, promoted posts, injected widgets).
- Detect feeds automatically (structural heuristics, optionally refined by an LLM), by inspecting an element, or from CSS/XPath selectors you type.
- Confirm/edit the proposals in a Shadow-DOM overlay, or programmatically with headless helpers.
-
Iterate a unified feed that keeps up with infinite scroll (
MutationObserver) or pagination (ingest), with a report per item:kept,excluded(and why), orpending. - Exclude items by specificity-ranked selector rules, built-in ad rules, or plain-language LLM criteria. Nothing is removed unless you call a removal helper.
It runs in browser extensions (Chromium + Firefox, MV3), headless browsers (Playwright/Puppeteer injected bundle), and server-side (Bun/Node with a parsed DOM), using the same API everywhere.
Install
…How I Built It
The open-source AI part
AI is optional and off by default. Ghostify is perfectly happy as a dumb "hide this post" button. Turn AI on, and it gets three jobs, each with its own model:
| Job | Default (local, via Ollama) | Used for |
|---|---|---|
| Embeddings | embeddinggemma-2 |
Semantic search over your ghosts |
| General | gemma4:e2b |
Summaries and tags for each ghosted post |
| Decision | d3:2b |
Checking new feed items against your auto-ghostify rules |
All three are open-weight and small enough to run on a laptop. Start Ollama so it accepts requests from extensions, and that's it:
ollama pull embeddinggemma-2
ollama pull gemma4:e2b
ollama pull d3:2b
OLLAMA_ORIGINS="chrome-extension://*,moz-extension://*" ollama serve
If you'd rather use a hosted model, every job can switch to OpenRouter, Gemini or OpenAI with your own key. The plumbing is the AI SDK, so swapping providers is a dropdown, not a rewrite.
A few things I'm happy about:
- Fallbacks everywhere. If the decision model is down, the general model handles the rule check. If embeddings fail, search falls back to keywords. Nothing breaks because Ollama isn't running.
- Embeddings are tagged with the model that made them. Change the embedding model and Ghostify ignores the old vectors instead of mixing them up. Reindex fills the gaps.
-
One honest gap:
d3's response format isn't documented (yet). Ghostify reads the shapes it knows, treats anything else as "no match", and logs it. Better to miss a ghost than to hide something you wanted to see!
The extension part
- WXT for one codebase that builds Chrome MV3 and Firefox, plus the new-tab flavor.
- React 19 + Tailwind v4 for the popup, the Ghostify page and settings.
- Dexie (IndexedDB) for ghosts, embeddings and keys. Everything stays in your browser. No account, no sync, no server.
-
Feedgrab for finding feeds and items, and for hiding them. Ghostify never touches a feed node itself; it asks Feedgrab's
removeItemto hide it, and its own UI lives in a separate Shadow DOM.
Security was a bit of a rabbit hole:
- Content scripts never see API keys. They don't even open IndexedDB. Everything goes through typed messages to the background script, and the background talks to the model.
- Keys aren't in
storage.localeither, because content scripts can read that. They live in a Dexie table on the extension's own origin. - In content scripts, Feedgrab gets a proxy instead of a model, which also shrank the bundle from 380 KB to 71 KB. Nice side effect!
The "how did you build all this in a week" part
I didn't write most of the code by hand. I used a coding agent.
But I didn't just say "build me an extension" either. I've been using a small spec-driven flow I call specx:
- Spec: I wrote the spec by hand. It's 17 lines. Literally a paragraph and a list of features.
- Plan: the agent turned it into a plan with the stack, architecture and a Decisions log.
- Tasks: then into 23 tasks across 4 phases, each with the spec/plan line ranges it needs, the files it touches, and a Done check.
- Build loop: take the first unblocked task, read only what it references, implement it, run its Done check plus lint + typecheck + tests, tick the box, commit. Repeat.
- Phase check: at the end of each phase, follow the README from a clean state, build, load it in the browser, fix what's broken, tag a release.
The rules live in AGENTS.md, so any agent can pick it up. Every judgment call goes into the plan's Decisions log, so I can see why something was done the way it was. And there's one commit per task, so git log reads like a changelog.
Phase 1 was "ghost a post on DEV". Phase 4 was "auto-ghostify with local AI". Each phase ended with something I could actually load and use, not a pile of half-wired modules.
Why Does Open Innovation Matter?
Because the thing Ghostify reads is my browsing.
Think about what an AI feature in a feed extension actually sees: every post on every feed you scroll, what you hide, what you save for later, what you search for. That's a pretty detailed picture of a person. I don't want that going to anyone's API by default. Not even mine.
With open-weight models, it doesn't have to:
- Gemma 4 E2B, EmbeddingGemma 2 and d3:2b run on my laptop. Summaries, tags, semantic search and the rule checks all happen locally. With Ollama, nothing leaves the machine.
- It costs nothing per post. Auto-ghostify checks every new item in the feed. With a paid API, that's a meter running every time I scroll. Locally, it's just my fan spinning a bit.
- It keeps working. No rate limits, no deprecation emails, no "this model will be retired on…".
- A small, specialized model is enough. I don't need a giant model to answer "is this a weekly discussion thread?" A 2B decision model is the right size for the job, and I could only pick it because it's open.
And it's not just the models. Ghostify is built on open source all the way down: WXT, React, Dexie, the AI SDK, Biome, Vitest… and Feedgrab, which I split out because I wanted it to be reusable. If someone wants to build a "hide the reruns" extension for YouTube, or a feed reader for a site that doesn't have RSS, the hard part is already there. MIT licensed. Go for it.
A closed API would have given me a smarter summary, maybe. Open models gave me a tool I'm actually comfortable running on everything I read.
My Agent Session
The whole project was built in agent sessions, following the specx flow above. The best way to see them is the commit history: one commit per task, from Scaffold WXT React project to Add auto-ghostify rules with rescue.
The Spirit of It
I keep thinking about the name.
Most "hide this" features are really "delete this". You click, it's gone, and you'll never find it again. That always felt a bit wrong to me. I did read that post. Some of them were good! I just don't need to see them every time I open the home page.
So in Ghostify, nothing dies. It becomes a ghost. It stays in the machine, with its title, its author, a little summary, and a few tags, waiting until you call it back. A bit like Día de los Muertos for posts: not here anymore, but not forgotten either.
And somewhere in an archived repo, myDEV is resting peacefully, knowing it started all of this. 🕯️
Prize Categories
-
Best Use of Gemma: Ghostify uses Gemma 4 E2B (
gemma4:e2b) for summaries and tags, and EmbeddingGemma 2 (embeddinggemma-2) for semantic search, both running locally through Ollama by default.
Thanks for reading! And if a post of yours disappears from my feed, don't take it personally. It just means I've probably already read it, seen it, or will read it later. 👻
Top comments (2)
Cool bro!
Let me know when it's ready to use! 😍
Thanks sis! 🥰
Will do ... 🙏