Khoj

What Khoj does

Khoj is a personal AI server built around your own material. It indexes notes and documents (Markdown, Org-mode, PDF, Word and plain text files, Notion pages and GitHub repositories) for semantic search, and its chat pulls the most relevant passages into each answer and shows which ones it used. Slash commands switch the behavior: /notes answers only from your files, /online searches the web, /research takes several steps for a more thorough answer, /code runs Python, and /image and /diagram create pictures or Excalidraw drawings. Agents bundle a persona, a knowledge set, a model and tools, and automations run a query on a schedule and email you the result, such as a weekly news digest.

You can reach the same server from the web app, a desktop app, an Obsidian plugin or an Emacs package, which makes it a natural fit for people who already keep their notes in those tools.

Models and providers

Hosted models come from OpenAI, Anthropic and Google, including through Vertex AI. For local models, Khoj talks to any OpenAI-compatible server: the docs cover Ollama, LM Studio, vLLM and a LiteLLM proxy, set up through OPENAI_BASE_URL or the admin panel. Some setup pages still describe running GGUF models inside Khoj with GPU build flags, but the current code no longer includes that engine, so plan on a separate model server. Embeddings for document search are computed locally with sentence-transformers models.

The docs warn that, with a hosted model, the passages Khoj retrieves from your notes are sent to that provider along with your question.

Privacy and data

Indexed content, embeddings and conversations are stored in your PostgreSQL database. Self-hosted servers send anonymized usage telemetry by default: which client and features are used, which content types are configured and basic request metadata, but not your content. The docs say it goes to PostHog. Set KHOJ_TELEMETRY_DISABLE=True to turn it off.

Self-hosting Khoj

The official Docker Compose file starts PostgreSQL with pgvector, the Khoj server on port 42110, SearXNG for web search, a Terrarium sandbox for code execution and a virtual desktop container used by its computer-operator feature. Before the first start you set an admin password and a Django secret key, plus API keys or a local model URL. A pip install (Python 3.10 to 3.12) can use an embedded Postgres instead.

The default setup runs in anonymous single-user mode. Removing that flag turns on accounts: magic-link sign-in by email (sent automatically with a Resend API key, or copied by hand from the admin panel) or Google OAuth, which needs the production image. Emailed automations also need Resend. For search, you can swap SearXNG for Serper, Firecrawl or Exa; E2B can replace the local code sandbox.

Hosted version

None any more. Khoj Cloud closed on April 15, 2026. The team said it would focus on other products and that Khoj stays open source for self-hosting. The README and parts of the docs still point to the cloud app.

Limitations

Development has slowed since the cloud shutdown: the last release, 2.0.0-beta.28, came out in March 2026, and recent commits are mostly fixes and dependency updates. The documentation is partly out of date. The stack is heavier than single-container chat apps, and telemetry is opt-out.

Who it suits

Khoj suits people who want an assistant that knows their own notes, especially Obsidian and Emacs users, and who are comfortable running Docker and Postgres. If you mainly want a multi-model chat UI for a team, LibreChat or AnythingLLM are closer to that job; for cited web answers alone, see Vane.