Lobe

Lobe turns deliberate X bookmarks into organized, searchable context. You keep using X's native bookmark button. Lobe captures the post at that moment, returns control immediately, then classifies why you saved it in a durable background job.

This repository contains the complete first product slice:

• A Chrome Manifest V3 extension built with WXT and React 19
• A Bun and Hono API with a durable background worker
• PostgreSQL, Drizzle, and pgvector persistence
• Intent classification and embeddings through Vercel AI SDK
• GPT-5.6 Luna for classification and selector repair
• text-embedding-3-small embeddings with indexed, sub-200 ms retrieval
• A responsive React 19 library with search, filters, corrections, and taste signals
• A shared shadcn component package, Hugeicons, Geist, and a pure-black theme

The current scope is intentionally explicit-save only, X only, and does not include MCP or passive browsing analysis.

How the X integration works

The extension does not redraw or replace X's bookmark icon. It listens for a click on X's real semantic controls:

article[data-testid="tweet"]
button[data-testid="bookmark"]
button[data-testid="removeBookmark"]

That keeps X responsible for the exact SVG, spacing, hover state, animation, and saved state. Lobe extracts only the nearest post, including quoted-post boundaries, media, author, canonical status URL, and an optional rendered screenshot.

The selectors live in a versioned recipe. When a layout no longer matches, the extension reports a compact semantic DOM sketch with no post text or handles. The server can derive one replacement recipe with GPT-5.6 Luna, validate it, cache it, and share it with every extension install.

Monorepo

apps/
  extension/   WXT extension, X capture, popup, and settings
  server/      Hono API and background worker
  web/         Searchable bookmark library and taste profile
packages/
  ai/          Vercel AI SDK classification, embeddings, and recipe repair
  db/          Drizzle schema, migration, repositories, and job queue
  shared/      Runtime schemas, intent metadata, messages, and recipes
  ui/          Shared shadcn primitives and Tailwind design tokens

The original product exploration is preserved in lobe-brainstorm-report.html.

Local setup

Requirements: Bun and Docker.

bun install
cp .env.example .env
openssl rand -hex 24

Put the generated value in LOBE_API_TOKEN. Add OPENAI_API_KEY to use live AI classification and semantic embeddings. The OpenAI key stays on the server and must never be put in a VITE_ or WXT_PUBLIC_ variable.

Start and migrate PostgreSQL:

bun run db:up
bun run db:migrate

Run the API and web app in separate terminals:

bun run dev:server
bun run dev:web

Open http://localhost:5173, then connect with the same LOBE_API_TOKEN.

Load the Chromium extension

Build the unpacked extension:

bun run --filter '@lobe/extension' build

Then:

1. Open chrome://extensions in Chrome, Brave, Arc, Edge, or another Chromium browser.
2. Enable Developer mode.
3. Choose Load unpacked.
4. Select apps/extension/.output/chrome-mv3.
5. Open Lobe's extension settings and enter http://localhost:8787 plus your LOBE_API_TOKEN.
6. Bookmark or unbookmark a post on X normally.

For live extension development, run bun run dev:extension instead.

Save pipeline

native X bookmark click
  -> persist capture and queue job in one transaction
  -> return 202 immediately
  -> classify intent in the worker
  -> create pgvector embedding
  -> mark ready and update the library

Intent is inferred as one of: Try it, Build similar, Learn, Reference, Buy, or Share. A confident result stays silent. A low-confidence result shows a compact, nonblocking prompt in the lower corner of X. The user can select the best intent and explain why the post mattered, or dismiss the prompt and leave the save marked as unsure. Submitted explanations are attached to the save and retrieved through the post's vector as examples when similar posts are classified later. Every unsure save also stays available in the library's Needs review queue.

Without an OpenAI key, the API remains usable with transparent deterministic classification and PostgreSQL text search. It never presents fallback output as an AI result.

Validation

Run the complete local gate:

bun run check

The database integration suites are explicit because they require the Docker database:

bun run --filter '@lobe/db' test:db
bun run --filter '@lobe/server' test:db

Benchmark the indexed search path with 10,000 generated saves and a 200 ms failure budget:

bun run benchmark:vector

The generated Chrome artifact is available under apps/extension/.output/chrome-mv3 after a build, or as a zip after bun run --filter '@lobe/extension' zip.