khoj-ai/khoj

A self-hostable personal AI application that answers questions from the web or your own documents, using local or hosted LLMs.

What it is

Khoj is a Python application that acts as a "second brain": it lets you chat with a language model that can retrieve answers from the internet and from your own files. It targets individuals who want a private, self-hosted AI assistant, as well as teams that need it as a cloud or on-premises service. The project spans on-device use up to cloud-scale deployment, and connects to a range of local and online models (llama3, qwen, gemma, mistral, gpt, claude, gemini, deepseek).

Key features

Chat with local or online LLMs, including offline models via llama.cpp/Ollama-style setups and hosted providers (OpenAI, Anthropic, Google).
Retrieval over your own documents, including images, PDF, Markdown, org-mode, Word, and Notion files.
Semantic search over indexed documents using sentence-transformer embeddings and pgvector.
Custom agents with their own knowledge, persona, chat model, and tools.
Scheduled automations that deliver newsletters and notifications (backed by APScheduler and cron scheduling).
Multiple client surfaces: web, Obsidian, Emacs, desktop, phone, and WhatsApp.
Additional capabilities noted in the README: image generation, voice/text-to-speech, and code execution (the manifest includes an e2b code interpreter dependency).

Tech stack

Python, required >=3.10, <3.13.
Web/API layer: FastAPI (>=0.110.0) served with Uvicorn/Gunicorn; websockets for streaming.
Application/data layer: Django (5.1.15) with django-unfold admin, PostgreSQL via psycopg2, and pgvector (0.2.4) for vector search; local Postgres via pgserver for the local extra.
ML/embeddings: sentence-transformers (3.4.1), transformers (>=4.53.0), torch (2.6.0), tiktoken.
LLM SDKs: openai (>=2.0.0, <3.0.0), anthropic (0.75.0), google-genai (1.52.0), plus the mcp package (>=1.23.0).
Document parsing: PyMuPDF, docx2txt, markdownify, markdown-it-py, BeautifulSoup, lxml, rapidocr-onnxruntime; openai-whisper for speech-to-text.
Scheduling/retrieval: APScheduler with django_apscheduler, langchain-text-splitters and langchain-community for chunking/retrieval helpers.
Auth and messaging: Authlib, phonenumbers; optional prod extras add Stripe, Twilio, and boto3.
Packaging: hatchling with hatch-vcs; tooling includes mypy, ruff, and pytest (pytest-django, pytest-asyncio).
Clients live in the repo tree as well, including an Android app (Java/Gradle) under src/interface.

When to reach for it

You want a private, self-hosted AI assistant that keeps your documents on your own infrastructure.
You need question answering grounded in a personal corpus spanning mixed formats (PDF, Markdown, org-mode, Word, Notion, images).
You want to use offline/local models rather than sending data to a hosted provider, or mix local and hosted models.
You want the same assistant reachable from several clients (web, Obsidian, Emacs, desktop, phone, WhatsApp).
You want to build custom agents and schedule recurring research or notifications.

When not to reach for it

You need a lightweight, embeddable library rather than a full application. Khoj is a Django + FastAPI service with PostgreSQL and pgvector as core dependencies, so it carries real deployment weight.
You cannot run or manage PostgreSQL/pgvector and a Python 3.10–3.12 environment with a torch-based ML stack.
AGPL-3.0 licensing is incompatible with your distribution or product model.
You only need a single-format, small-scale search and don't want the overhead of embeddings, a database, and multiple model integrations.

Maturity signal

Khoj is actively maintained. It carries roughly 35.9k stars, was created in 2021, and had a recent push (mid-2026 in the provided facts), with the pyproject classifiers marking it "Production/Stable". The dependency list pins recent versions across the ML and web stack and integrates current model SDKs, which indicates ongoing upkeep rather than long-tail maintenance. A relatively low open-issue count (123) against the star count and an active contributor base and Discord suggest a project with steady development and community support.

Alternatives

Open WebUI — reach for it when you primarily want a polished chat front end over Ollama/OpenAI-compatible models and don't need Khoj's document-corpus and multi-client focus.
AnythingLLM — reach for it when you want a document-chat/RAG desktop or server app with a simpler stack and less emphasis on editor/messaging clients.
PrivateGPT — reach for it when your main goal is fully local document Q&A and you prefer a smaller, more focused codebase over Khoj's broader feature surface.

Notes

Although the README frames Khoj as a personal/local AI, the codebase is built on Django and PostgreSQL with pgvector, closer to a server application than a desktop tool.
The repository bundles native clients directly, including an Android app (Java/Gradle) alongside the Python backend.
Optional prod extras pull in Stripe, Twilio, and boto3, reflecting the hosted cloud/enterprise offering described in the README rather than the self-hosted path.
The license is AGPL-3.0, which is notable for anyone considering deploying it as a network service or building a derived product.

Wiki: khoj-ai/khoj