Ghostify

Ghostify is a browser extension (Chrome and Firefox) for feeds that keep showing you the same posts, or posts you probably want to deal with later. Hover a post, click the ghost, and it disappears from the site's feed. Ghostify keeps a copy so you can find them again later, snooze them, read later, run AI models on them, and many more interesting features are coming up!

Latest: v1.5.3.

• In-browser AI models keep working after one fails to load (v1.5.2).
• Sort your ghosts by date, relevance or at random, ascending or descending; the choice is remembered (v1.5.1).
• Choose which link identifies a post (x.com posts keep their own URL), and check selector edits against a saved sample of each feed (v1.5).
• Edit any site's page selectors by hand, and restore built-in sites to how Ghostify ships them (v1.4.5).
• Move the ghost button to any corner of a post, for all sites or per site (v1.4).
• dev.to works logged in on every home tab (v1.3).
• AI runs in your browser by default, with no GPU, Ollama or API key needed (v1.2).
• Read full posts and ask AI about them in the viewer; every AI job can be paused or stopped (v1.1).

See CHANGELOG.md for every release.

• The dev.to home feeds, logged in or out (Relevant /, Latest /latest, Top /top/*, and when logged in also /discover, /following, /following/latest, /curated and /latest_less_filtered) 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. If it covers something on a site, move it in Settings → Appearance to any corner of the post, with its distance from the edges in px, for all sites or just one.

• Ghosted posts stay hidden when you reload, and also when they come back later through infinite scroll.

• The popup lists your recent ghosts. From there you can open one, unghost it so it shows on the site again, or open the full Ghostify page.

• The Ghostify page has three tabs:

  • Ghosted is everything you hid.
  • Later is what you want to read.
  • Snoozed holds posts that come back later.

• Snooze a post until later today, tomorrow, the weekend, next week, or a date you choose. The post stays hidden on the site, and when the time is up it shows in Back from snooze at the top of Ghosted.

• Search your ghosts by keyword and filter them by site. Sort them by date ghostified or relevance (to the search; by date without one), descending, ascending or random (Shuffle again reshuffles); the choice is remembered. Long lists load 50 at a time (Show more), and Unghost on the Ghostify page has an Undo.

• The viewer opens when you click a post's title, thumbnail or excerpt. It has three tabs:

  • Overview: the summary, tags and excerpt.
  • Reader: Load full content fetches the post and shows its main text inside Ghostify, without the site's ads and menus (the first time, you're asked to allow access to that site). Reload fetches it again.
  • Ask AI: Summarize the full content, or ask your own questions about it. Answers stream in as they're written, and each one can be paused, resumed or stopped.

  Toolbar buttons open the original post in a new tab or a new window, open the viewer as its own page (which you can also move to a new window), or widen the panel. The popup can open Ghostify in a new window.

• Add any site. On a page with a feed, open the popup and click Add this site. Ghostify asks for access to that site only, and then:

  1. feedgrab proposes the feeds on the page, and its overlay lets you confirm them, pick containers and items, or type CSS/XPath selectors.
  2. You choose which page paths the feeds apply to.
  3. You choose whether they're for the logged-in, the logged-out or any view. For a specific view, you can pick an element that only shows in it (for example your avatar), so Ghostify can tell the views apart.

  Edit selectors in the popup reopens the overlay on a site you already added.

• Settings → Sites lists your sites with a search box, and each site opens on its own page (with a breadcrumb back to the list). It lets you:

  • turn sites on or off and delete the ones you added;
  • see and edit each page's path pattern, view and selectors (edited built-in selectors are kept over Ghostify's updates);
  • restore a built-in site, or one edited built-in page, to how Ghostify ships it;
  • edit the login indicator;
  • export and import your sites as JSON.

  dev.to is built in. Its login indicator is #nav-profile-image, and its home feeds are set up for both the logged-in and logged-out views.

• The view switches as soon as you log in or out, without a reload.

• The new-tab flavor (ghostify-newtab-* zips) puts the Ghostify page on every new tab. Install either the regular flavor or the new-tab flavor, not both.

• AI is optional and off by default. Turn it on in Settings → AI and download the in-browser models (or pick another provider) to get:

  • summaries and tags for each ghosted post, written by the general model. Click a tag to filter by it.
  • semantic search: turn on Semantic next to the search box and describe what you remember.
  • summaries of the full content and answers to your questions in the viewer's Ask AI tab;
  • auto-ghostify rules: in Settings → Sites, write plain-language rules per site, such as "crypto hype" or "posts that are mostly self-promotion". The decision model checks new feed items against them and ghosts the matches. Auto-ghosted posts are labelled with their rule. Not this brings a post back, and that rule never hides it again.

• AI activity (the pulse button in the Ghostify page header, or Settings → AI activity) lists every AI request: summaries, tags, indexing, rule checks, searches, model tests and questions. Expand a row to see its model, prompt, input, output and error. Each job can be paused, resumed or stopped, and Pause all, Resume all and Stop all act on everything at once. A pause lasts across browser restarts. While AI is paused, searches fall back to keywords and auto-ghostify rules don't run.

• Settings → Prompts lets you change what Ghostify asks the AI: the summary-and-tags prompt, the Summarize and Ask prompts, and the rule question sent to System One decision models. Reset to default restores the original prompt.

• Everything stays local in your browser (IndexedDB), with no account and no sync. With in-browser models or Ollama, AI stays on your machine too.

Each purpose (embeddings, general model, decision model) can use its own provider and model.

Small open models run inside the browser with Transformers.js (ONNX Runtime Web). They need no GPU, no Ollama and no API key. No model ships with the extension and nothing downloads on its own: you choose what to download.

1. Open Settings → AI and turn on Use AI.
2. Click Download on each row, or Download all (about 460 MB for the defaults). The files come from Hugging Face once and stay in the browser's cache. Progress shows on each row, and downloads keep going if you close the settings page.
3. Click Test on each row. Delete removes a model's files.
4. Optional: under In-browser models, pick GPU (WebGPU) instead of CPU if your browser supports WebGPU. It is faster, but uses different model files, so download again.

Any Hugging Face ONNX repo works too: choose Other… in the model list and type its id (owner/name).

How it works and what to expect:

• Models run in a worker on an offscreen page (Chrome) or a hidden frame of the background page (Firefox). They stop after 10 idle minutes to free memory.
• A model needs roughly 10–15 times its file size in RAM while it runs. On a computer with 4 GB of RAM, choose SmolLM2 135M.
• Summaries take a few seconds per post on a laptop CPU. The viewer reads at most 6,000 characters of an article with in-browser models (24,000 with others).
• Rule checks use a zero-shot classifier ("This item is about …"), not a chat model. Short topics such as "crypto hype" or "an advertisement" work much better than long descriptions. The rule question in Settings → Prompts applies only to System One models.

Any Ollama model with the decision capability (for example tev1:0.8b or tev1:4b, or d1-omni) works as the decision model. So does an OpenRouter decision model such as Jev (typesafe/jev-1.14). Ghostify sends these models System One requests (/v1/systemone); any other model gets a chat prompt.

1. Pull the models:

   ollama pull embeddinggemma-2
   ollama pull gemma4:e2b
   ollama pull d3:2b

2. Let Ollama accept requests from browser extensions. Without this step, the browser console shows a CORS error. Set OLLAMA_ORIGINS on the Ollama server that is actually running. If you changed the Ollama URL in Settings, set OLLAMA_HOST to the same host and port.

   • Linux: the installer runs Ollama as a systemd service, so setting the variable in a terminal has no effect. Run sudo systemctl edit ollama and add:

     [Service]
     Environment="OLLAMA_HOST=127.0.0.1:11434"
     Environment="OLLAMA_ORIGINS=chrome-extension://*,moz-extension://*"

     Then run sudo systemctl restart ollama.
   • macOS: run the commands below, then quit and reopen the Ollama app:

     launchctl setenv OLLAMA_HOST "127.0.0.1:11434"
     launchctl setenv OLLAMA_ORIGINS "chrome-extension://*,moz-extension://*"

   • Windows: add OLLAMA_HOST and OLLAMA_ORIGINS as user environment variables, then quit Ollama from the tray and start it again.
   • No background service: run

     OLLAMA_HOST=127.0.0.1:11434 OLLAMA_ORIGINS="chrome-extension://*,moz-extension://*" ollama serve

   To check, run curl -i -H "Origin: chrome-extension://x" http://127.0.0.1:11434/api/tags. A 403 response means the running server didn't pick up OLLAMA_ORIGINS.

3. In Settings → AI:

   1. Turn on Use AI.
   2. Click Test on each row.
   3. Wait for Search index to fill. Ghosts saved before AI was on, or before you changed the embedding model, are embedded in the background, one at a time, whenever no other AI work is waiting. The card shows how many are done, the estimated time left, and why indexing paused (for example, Ollama isn't running). It retries every 5 minutes.

1. Pick OpenRouter, Gemini or OpenAI for any purpose. The model box suggests models, and you can type any model id.
2. Paste the key into API keys.

Keys are stored only in the extension's own storage and sent only to that provider. Content scripts on the sites you visit never see them.

If the decision model fails (for example, Ollama isn't running), the general model handles the rule check instead. If embeddings fail, search falls back to keywords.

1. Download a zip from Releases:
   • ghostify-<version>-chrome.zip for Chrome, Edge or Brave;
   • ghostify-<version>-firefox.zip for Firefox;
   • ghostify-newtab-<version>-chrome.zip or ghostify-newtab-<version>-firefox.zip if you want Ghostify on every new tab.
2. Chrome:
   1. Unzip the file.
   2. Open chrome://extensions and turn on Developer mode.
   3. Click Load unpacked and select the unzipped folder.
3. Firefox:
   1. Open about:debugging#/runtime/this-firefox.
   2. Click Load Temporary Add-on… and select the zip file, or its manifest.json after unzipping.
   3. Firefox removes temporary add-ons when it restarts.

Requires Bun 1.3 or later.

bun install
bun run build           # Chrome  → .output/chrome-mv3
bun run build:firefox   # Firefox → .output/firefox-mv3

# New-tab flavor (optional)
bun run build:newtab           # → .output/chrome-mv3-newtab
bun run build:newtab:firefox   # → .output/firefox-mv3-newtab

Load the build unpacked:

• Chrome: on chrome://extensions, turn on Developer mode, click Load unpacked, and select .output/chrome-mv3.
• Firefox: on about:debugging#/runtime/this-firefox, click Load Temporary Add-on… and select .output/firefox-mv3/manifest.json.

For the new-tab flavor, load .output/chrome-mv3-newtab or .output/firefox-mv3-newtab/manifest.json the same way.

Then:

1. Open https://dev.to/, hover a post, and click the ghost.
2. Click the Ghostify toolbar icon, then Open Ghostify. If you use the new-tab flavor, open a new tab instead.
3. To try another site, open it, click the toolbar icon and choose Add this site.
4. Optional: set up AI (see AI setup).

Pushing a v* tag runs .github/workflows/release.yml. It tests the extension, builds the Chrome, Firefox and Firefox sources zips for both flavors, and attaches them to a GitHub Release:

git push origin v1.0.0

• The spec is specs/SPEC-v1.md, the plan is specs/PLAN-v1.md, and the tasks are specs/TASKS-v1.md.
• This version covers every task, T01–T31 (Phases 1–5), and all of the spec:
  • feedgrab feed reading;
  • the dev.to home feeds (logged in and out), plus adding more sites;
  • custom feed selectors through feedgrab;
  • separate logged-in and logged-out views with their own selectors;
  • the ghost icon on recognized items;
  • removing ghosted items from the feed and keeping them in the extension;
  • viewing later, snoozing, and the Ghostified tab until you remove an item;
  • keyword filtering;
  • the UI;
  • the new-tab option (as a build flavor);
  • local Ollama models (embeddinggemma-2, Gemma 4 E2B, d3:2b) and bring-your-own-key OpenRouter, Gemini and OpenAI, configurable per purpose.
• Phase 5 (T24–T31) came out of a security, performance and UI audit, and added in-browser models:
  • security: the background accepts only the feed messages from web pages, and only about that page's site. Saved posts keep only http(s) links, and their text is capped. Imported sites are validated. Full-content loads are limited to 5 MB of HTML within 20 s. Model lists time out. The extension's content security policy allows no remote code;
  • performance: indexed, limited item lists; search reads only the current model's vectors; a single backlog per indexing run; fewer content-script restarts; a throttled login-view watcher; and a ghost button placed once per frame;
  • UI: Undo for unghosting, paged lists, a viewer that stays open when its post moves tabs, settings that stay in sync across pages, focus handling for drawers, and a confirmation before deleting a site;
  • in-browser models, which are now the default AI provider (see AI setup).

• Ghostify isn't published to the Chrome Web Store or Firefox Add-ons. Install it from a release zip or a local build.
• Decision models get one yes/no (noul) System One question per rule. Ghostify reads the documented { "noul": <probability> } answer and a few lenient variants (see src/ai/decision.ts). It treats anything else as "no match" and logs it.
• Load full content shows what the page sends to Ghostify, using your cookies for that site. Pages built entirely in the browser by scripts may have little readable text. In that case, open the original.
• Pausing a job that is already running stops its request, and Resume starts it again from the beginning. Rule checks and searches can't be paused, because a page is waiting for them, so pausing them stops them.
• Data isn't synced between browsers. To move your sites, use export and import in Settings → Sites.
• In-browser models are small: summaries are plainer than those from larger models, and the smallest chat models sometimes leave out tags. They run on one CPU thread, because extension pages aren't cross-origin isolated. The extension includes ONNX Runtime's WebAssembly (about 27 MB).

MIT

The extension icon is the 👻 glyph from Noto Emoji (Apache-2.0).