librarian

License: source-available, not open source. Installing from PyPI doesn't change the terms. librarian is free under PolyForm Noncommercial 1.0.0 (personal, hobby, research, education, nonprofits, government) or PolyForm Small Business 1.0.0 (fewer than 100 people and under about $1.3 million in revenue). Every other business needs a commercial license. librarian init asks which one covers you and keeps the answer on your machine. Full terms: LICENSE.md.

A local, private knowledge library for AI coding agents. librarian ingests PDFs, web pages, whole websites, code, markdown, images, and videos, keeps git repositories, Obsidian vaults, Zotero, feeds, and bookmarks in sync, indexes it all with hybrid semantic and keyword search in one SQLite file, and serves it to your coding agent through an MCP server, a REST API, and a CLI. Indexing, embedding, and search run on your machine: no cloud vector database and no API keys.

The distribution is librarian-toolkit (the PyPI name librarian belongs to an unrelated project). The command and the import name are librarian.

Install

Python 3.11 or newer, on Windows, macOS, or Linux.

pip install librarian-toolkit                 # or: uv tool install librarian-toolkit
pip install "librarian-toolkit[gpu]"          # with an NVIDIA GPU (CUDA 12 runtime)
pip install "librarian-toolkit[media,browser]"  # video/audio, headless-browser fallback

Then, in the folder that will hold your library (library.db, models/, and .env are all relative to it):

librarian models download      # the embedding model, about 133 MB
librarian init                 # creates library.db and asks which license applies
librarian ingest https://docs.python.org/3/tutorial/classes.html --wait
librarian search "how do python class variables differ from instance variables"
librarian doctor               # checks the install

Connect it to an MCP client, for example Claude Code. -C names the library's folder, since the client starts the server in its own:

claude mcp add librarian -- librarian -C /path/to/your/library mcp

With media on an x86-64 machine, use the GPU build of ONNX Runtime alone. The media packages also pull in the CPU build, which can shadow the GPU one; librarian doctor warns when both are installed. The fix is pip uninstall onnxruntime, then pip install --force-reinstall --no-deps onnxruntime-gpu. A source checkout installed with uv sync avoids this.

More extras: research (the SearXNG search backend for the research loop), scholarly, brave, uspto (more search backends), encryption (SQLCipher), cloud (the Anthropic provider for research roles you allow off the machine). The versions that are tested together are pinned in the repository's uv.lock; a source install with uv sync uses exactly those.

Documentation

• Installation
• User guide
• MCP server
• Configuration
• Changelog
• Security policy and support policy