AI auto-tagging
LLMs read every item, assign tags with confidence scores, and generate 2–3 sentence summaries automatically. Bring your own model - Ollama, Groq, NVIDIA, or any OpenAI-compatible endpoint.
Reach into your pocket,
pull out exactly what you need.
4dpocket is a self-hosted AI knowledge base that saves, tags, summarizes, and connects everything you read - across 17+ platforms. Local-first. Private by default. Your data never leaves your machine.
A knowledge base that thinks
Every item is extracted, enriched, indexed, and connected - automatically. You just paste URLs.
Four search modes
Full-text (SQLite FTS5 + BM25), fuzzy fallback for typos, semantic similarity (ChromaDB vectors), and hybrid RRF fusion. Inline filter syntax like tag:devops is:favorite after:2024-01.
17 platform processors
Deep extraction for YouTube, Reddit, GitHub, X, Instagram, HackerNews, Stack Overflow, TikTok, Mastodon, Threads, Substack, Medium, LinkedIn, Spotify, plus PDFs and images. Paste a URL - we do the rest.
Self-hosted & private
Runs entirely on your machine. Docker, pip, or source. Per-user data isolation, SSRF protection on every processor, DOMPurify on all HTML paths, prompt-injection filtering before the LLM.
Collaborative sharing
Share items, collections, or tag groups with specific users. Viewer or editor roles. Public links with optional expiry. Knowledge feed to follow other users. Comments and discussions inline.
PWA + Chrome extension
Installable Progressive Web App with share target, offline access, and mobile-optimized nav. Chrome extension for one-click save, right-click context menu, and text highlight capture on any page.
17 platforms. One pocket.
Paste a URL. 4dpocket detects the platform, deeply extracts the content, and enriches it with AI.
Generic URL
YouTube
Reddit
GitHub
X / Twitter
Instagram
Hacker News
Stack Overflow
TikTok
Mastodon
Threads
Substack
Medium
LinkedIn
Spotify
Images (OCR)
PDF
Plus notes, highlights, RSS feeds, Chrome-bookmark imports, and Pocket exports - all unified under one search.
Three steps. Zero friction.
Save anything in one or two clicks. AI handles the organization. You just search when you need it.
Paste a URL. Or use the extension.
Drop any URL into the app, share from your phone, or one-click with the Chrome extension. We detect the platform and deeply extract title, content, media, comments, metadata.
AI tags, summarizes, connects.
Your preferred LLM assigns tags with confidence scores, generates a 2–3 sentence summary, finds related items via semantic similarity, and suggests the right collection.
Instant search across everything.
Full-text, fuzzy, semantic, or hybrid. Typo-forgiving. Filter by tags, dates, sources with inline syntax. Search items and notes together. Press Cmd+K from anywhere.
Bring your own model.
Run a local LLM for true privacy, or plug in your favorite cloud provider. No vendor lock-in.
Ollama (local)
Groq
NVIDIA
OpenAI-compatible
Anthropic-compatible
Admin panel overrides .env at runtime - change model, API key, or endpoint without a restart. No litellm, no langchain, no vendor SDK churn.
Up and running in 60 seconds.
Zero config. Zero external services. SQLite by default. Scale up when you need to.
Docker (recommended)
Single container, SQLite, no other services. Just works.
$docker run -d -p 4040:4040 \
-v 4dp-data:/data \
ghcr.io/onllm-dev/4dpocket:latest
Python package
Install from PyPI, run the interactive setup wizard.
$pip install 4dpocket
$4dpocket setup
$4dpocket start --sqlite
Open http://localhost:4040 - no login needed in single-user mode. See the Quick Start guide for PostgreSQL, Meilisearch, Ollama, and ChromaDB profiles.
Questions, answered.
What is 4dpocket, exactly?
4dpocket is a self-hosted, AI-powered personal knowledge base. It's not a bookmark manager - it extracts full content from 17+ platforms, enriches it with AI (tags, summaries, connections), and makes everything instantly searchable. Think of it as Pocket + Raindrop + Notion, but local-first and actually smart.
Is my data really private?
Yes. 4dpocket runs entirely on your machine - via Docker, pip, or source. There is no SaaS backend, no telemetry, no phone-home. If you pair it with Ollama for local LLMs, your content never touches the internet. Per-user isolation is enforced at every query (
WHERE user_id = current_user.id).
Do I need a GPU?
No. You can run 4dpocket against any cloud provider (Groq, NVIDIA NIM, OpenAI-compatible, Anthropic-compatible) with zero local compute requirements. If you want full privacy, Ollama runs small models comfortably on a modern laptop (8–16GB RAM). A GPU just makes local LLMs faster.
How much does it cost?
Free. 4dpocket is open source under GPLv3. Your only costs are your server (which can be
localhost) and optionally your AI provider (Groq has a generous free tier; Ollama is free forever).
What AI providers are supported?
Out of the box: Ollama (local), Groq, NVIDIA NIM, and any OpenAI-compatible or Anthropic-compatible endpoint. That covers essentially every modern LLM API. Switch providers at runtime from the admin panel - no restart required.
Can it work offline?
Yes. Pair it with Ollama and a local embedding model and 4dpocket runs completely offline - no network required beyond fetching the content you paste in. The PWA even caches static assets so the frontend loads without a connection.
How is this different from Pocket, Raindrop, or Notion?
Pocket and Raindrop are bookmark managers - they save links, not knowledge. Notion is a document tool - it's not designed for ingesting the firehose of things you read. 4dpocket sits in between: it captures content automatically, organizes it with AI, and is optimized for retrieval. And unlike all three, it's self-hosted and open source.
Does it support multi-user / teams?
Yes. Set
FDP_AUTH__MODE=multi and you get JWT auth, per-user isolation, collaborative sharing with viewer/editor roles, public links, and a knowledge feed. The first registered user automatically becomes admin.
Stop losing what you read.
Self-host in 60 seconds. Keep your knowledge where it belongs - with you.