# Backpack

> Backpack is an open-source, self-hostable multi-model AI research and
> chat app by Ayan Parkar: 6+ LLM providers, RAG research spaces,
> cross-chat memory, MCP servers and a versioned artifact workspace.

[← AYAN.WTF](https://www.ayn.wtf/) \ PROJECTS \ BACKPACK

## Overview

Backpack is an open-source AI research and chat app you can host
yourself. It talks to models from six providers, keeps your sources in
research spaces it can search, remembers things across chats, and writes
long documents into a versioned workspace beside the conversation.

- TYPE: Open-source AI research and chat app
- STACK: TypeScript, Postgres, Redis, RAG, MCP
- YEARS: 2024 to 2025
- REPO: [metaloozee/backpack](https://github.com/metaloozee/backpack)

![Backpack chat on the left and a versioned text artifact, a Rust lifetimes article, open on the right](https://www.ayn.wtf/projects/backpack-artifact-preview-full.webp)

The artifact workspace.

## Why I built it

I built Backpack to survive jumping into second-year engineering with a
month until finals. It was the hardest thing I'd shipped yet.

It grew into an open-source NotebookLM-style tool with web and academic
search, plus DeepResearch via
[rsrch.tech](https://github.com/metaloozee/rsrch.tech).

## Models

Backpack isn't tied to one vendor. The model picker covers Google,
Anthropic, OpenAI, Groq, OpenRouter and Cloudflare Workers AI, and you
only need credentials for the ones you turn on.

## Spaces and knowledge

A Space is a research workspace for one topic or project. It groups your
chats with the documents and web pages you add to it.

When you add a source, a background job pulls out the text, running PDFs
through OCR. It splits the text into 250-word chunks, embeds them and
stores the vectors in Postgres. The assistant then searches your
material and answers from it.

## Memory and tools

A Memory is a fact or preference Backpack saves and recalls in later
chats, so you don't repeat yourself in every new conversation.

The assistant has tools for web search, academic search, finance search,
page extraction and searching your own knowledge. You can also plug in
your own MCP servers, and their tools show up alongside the built-in
ones.

## Artifacts

Long answers don't belong in a chat bubble. Backpack writes them into an
artifact, a document that opens next to the conversation.

Every save is an immutable version. You can edit the text by hand, ask
the assistant for a targeted edit or a full rewrite, compare two
versions and restore an old one.

## Running it yourself

You need Bun, a Postgres-compatible database and Redis. Copy the example
env file, fill in the providers you want, run the migrations and start
the dev server.

## More projects

- [DITTO](https://www.ayn.wtf/projects/ditto): A durable AI coding
  workspace on the web. Each GitHub repo gets its own sandbox with
  parallel sessions: isolated git worktrees, resumable agent runs over
  SSE, and a real commit-to-PR workflow shared by agents and humans.
  Still in active development.
- [RANKMESH](https://www.ayn.wtf/projects/rankmesh): An autonomous SEO
  and AI-search visibility engine: agents audit, fix, write and publish
  while people review and approve. Multi-service FastAPI backend,
  Celery/Redis jobs with Claude and Codex runners, and a Next.js panel.
  This is what I ship day to day.
- [QUICKVID](https://www.ayn.wtf/projects/quickvid): Summarise, verify
  and chat with any YouTube video using transcripts, Gemini and RAG.
  Featured by LangChain, with 174 GitHub stars and 100+ active users
  before YouTube's crawler blocking made it unsustainable.

[← BACK TO AYAN.WTF](https://www.ayn.wtf/)

---
Screen font: Px437 IBM VGA 8x16, Ultimate Oldschool PC Font Pack
(int10h.org), CC BY-SA 4.0.