Khoj

Your AI second brain, honestly reviewed. RAG over your own docs, web search, agents — self-hostable under AGPL-3.0.

TL;DR

• What it is: Open-source (AGPL-3.0) personal AI assistant that indexes your documents and the web, then lets you chat with them, build custom agents, and schedule automations — think NotebookLM meets ChatGPT, but on your own server [3].
• Who it’s for: Researchers, knowledge workers, and founders drowning in documents who want semantic search across their own files without sending everything to Google’s cloud. Also useful for anyone who wants to swap AI models freely without re-building their knowledge base [2][3].
• Cost savings: ChatGPT Plus runs $20/mo with no document ingestion at scale and no self-hosting. Perplexity Pro is $20/mo. Khoj self-hosted runs on a modest VPS for under $10/mo with unlimited queries, your own models, and no data leaving your network [1][3].
• Key strength: RAG (retrieval-augmented generation) across PDFs, Markdown, Notion, Obsidian, Word docs, GitHub repos, and the live web — with citations showing exactly which source it pulled from [2][3].
• Key weakness: AGPL-3.0 license means embedding it in a commercial product requires open-sourcing your code or negotiating a commercial license. Setup is not beginner-friendly. Specific cloud pricing tiers are not publicly detailed [1][3].

---

What is Khoj

Khoj is a personal AI application built to extend your thinking rather than replace it. The core idea is simple: you point it at your documents and it creates semantic embeddings of everything, then answers your questions by retrieving the relevant chunks first before generating a response. That approach — called RAG — is what separates it from a plain chatbot and puts it in the same category as Google’s NotebookLM.

The difference from NotebookLM is that Khoj runs on your infrastructure, connects to whichever LLM you prefer (GPT-4o, Claude, Gemini, Llama, Qwen, Mistral, Deepseek — the README lists them all), and adds features NotebookLM doesn’t have: real-time web search, custom agents with specific personas, email automations, and image generation [README][3].

The GitHub description lands better than the homepage does: “Your AI second brain. Self-hostable. Get answers from the web or your docs. Build custom agents, schedule automations, do deep research.” The homepage, by contrast, has pivoted toward describing the company’s broader product portfolio — including Open Paper (a research workbench for academic papers) and Pipali (an on-device AI co-worker) — which makes it less obvious what the core Khoj app actually does on first visit [website].

The project sits at 33,463 GitHub stars and is backed by YCombinator [1]. It is licensed under AGPL-3.0 — not MIT, not Apache-2.0. That distinction matters and we’ll come back to it.

---

Why people choose it

The reviews cluster around one comparison: Khoj versus NotebookLM. The XDA Developers piece [3] — the most detailed hands-on review in the source set — frames it directly: “Enter Khoj, the first open-source tool I’ve found that closely matches NotebookLM’s functionality.”

What specifically beats NotebookLM according to that review:

Web search is on by default. NotebookLM added web search but requires deliberate interaction to trigger it. Khoj uses it automatically when your personal knowledge base doesn’t have an answer and cites its sources either way [3].

Chat history is more navigable. NotebookLM organizes around notebooks, which creates friction when you’re jumping between topics. Khoj lists all chats in a left-panel sidebar similar to ChatGPT, which works better for multi-subject workflows [3].

Global file search. You can search across all uploaded files instantly, not just the contents of a specific notebook [3].

Model flexibility. You’re not locked into Google’s model. If you want to run Llama 3 locally for privacy or Claude for quality, you swap the backend without rebuilding your knowledge base [2][README].

Agents and automations. NotebookLM has neither. Khoj lets you create agents with specific personas (a “teacher” agent that works through your study notes, a “copywriter” agent that knows your brand docs), and schedule automations that deliver newsletters or research summaries to your inbox [2][3].

The FutureTools entry [1] adds the offline angle: Khoj can run without internet access using a local LLM, which matters for anyone working with confidential documents who doesn’t want inference calls going to an external API.

The honest caveat from [3]: this comparison flatters Khoj relative to NotebookLM’s polish. The XDA review notes that Khoj’s layout is “much more manageable and approachable” but doesn’t claim it’s more refined. NotebookLM’s audio overviews and guided exploration features have no equivalent in Khoj.

On the benchmark front, the Khoj team published their own evaluation results [5] — which earns points for transparency even if you discount self-reported numbers. Running against Google’s FRAMES benchmark (multi-hop reasoning, temporal reasoning, tabular data) and OpenAI’s SimpleQA in Research mode, Khoj performed competitively. The evaluation harness runs automatically on every release via GitHub Actions, so the numbers are at least consistently measured rather than cherry-picked [5].

---

Features

Document ingestion and semantic search:

• PDFs, Markdown, org-mode, plain text, Word documents [README]
• Notion and Obsidian integrations [README][3]
• GitHub repository indexing [3]
• YouTube video content (via transcripts) [3]
• Semantic search using embeddings — matches meaning, not just keywords [3]

Chat modes:

• General (no retrieval — model’s base knowledge only)
• Default (single-shot retrieval from your docs and/or the web)
• Research (iterative multi-step retrieval, deeper reasoning, beta) [5]

Web search: On by default when your knowledge base is thin. Sources cited inline [3].

Custom agents: Create agents with specific knowledge bases, AI models, personas, and tools. Documented on the Khoj blog with step-by-step guides [README].

Automations: Schedule tasks — daily news summaries, research newsletters, event-triggered notifications delivered to your inbox [2][README].

Image and diagram generation: Generate images, charts, and visualizations directly from queries [2].

Voice interaction: Voice input and text-to-speech playback [README][1].

Multi-model support: Plug in GPT-4o, Claude, Gemini, Llama, Qwen, Mistral, Deepseek, or any HuggingFace model [1][2][README].

Platform access: Web app, iOS, Android, Obsidian plugin, Emacs plugin, WhatsApp integration [README].

AI explainability: Khoj shows which sources it pulled from and how it reasoned to reach an answer — this is the same RAG transparency that makes NotebookLM useful [2][3].

---

Pricing: SaaS vs self-hosted math

Khoj Cloud (app.khoj.dev): The free tier is accessible without setup — you can try it immediately at https://app.khoj.dev [README]. The merged profile data lists no public pricing tiers beyond “free,” and the website doesn’t surface paid plan details clearly. The FutureTools entry [1] mentions “special pricing for students.” Beyond that, specific monthly prices for paid cloud plans are not available in the source data — this review will not fabricate numbers.

Enterprise: Custom pricing, contact sales [website].

Self-hosted (Community Edition):

• Software license: $0 (AGPL-3.0)
• VPS: $5–15/mo depending on document volume and model inference
• LLM API costs: pay-as-you-go to OpenAI/Anthropic/Google, or $0 with a local model (Ollama)
• Your time to configure it

Comparison context:

• ChatGPT Plus: $20/mo — no doc ingestion at scale, no self-hosting, no custom agents
• Perplexity Pro: $20/mo — web search only, no personal document RAG, no self-hosting
• NotebookLM: currently free — but Google’s cloud, limited document types, no agents, no automations
• Self-hosted Khoj on Hetzner CX22 (~$6/mo): unlimited queries, full feature set, your models, your data

The math is straightforward for anyone spending $20/mo on ChatGPT Plus primarily to analyze documents: self-hosted Khoj replaces that bill with a VPS cost and one afternoon of setup.

---

Deployment reality check

The self-hosting path is Docker-based and documented at docs.khoj.dev. The README is the clearest install pointer.

What you need:

• A Linux VPS (minimum 4GB RAM if you’re running an API-backed LLM; significantly more if you want local inference)
• Docker and docker-compose
• PostgreSQL with the pgvector extension (for semantic search — this is the dependency that trips most users)
• An API key for whichever LLM you want to use, OR a local Ollama setup
• A reverse proxy (Nginx or Caddy) if you want HTTPS

What makes this harder than average: The pgvector extension for PostgreSQL is not a standard package on all distributions and cloud databases. It requires either a Postgres instance compiled with pgvector support or a managed database that includes it (Supabase, certain tiers of Neon, or RDS with the extension enabled). For non-technical founders, this is the most likely first-day blocker.

Connecting a local LLM for fully private inference (no API calls leaving your server) requires setting up Ollama separately — Khoj doesn’t bundle it. If you want the full privacy story — documents indexed locally, inference running locally, zero data to third-party APIs — you’re managing two separate services [1].

The XDA review [3] describes the experience as approachable for people comfortable with self-hosted tools but doesn’t claim it’s zero-configuration. For non-technical users, realistic time estimates: 2–4 hours for a working cloud-LLM-backed instance following the docs, potentially longer if pgvector setup has friction.

The AGPL-3.0 license deserves a specific callout here. AGPL is “copyleft with network use” — if you modify Khoj and let anyone use it over a network, you have to publish your modifications under AGPL too. For internal use (your own team, your own documents), this doesn’t matter. For anyone building a product on top of Khoj and offering it to customers, it does. The Khoj website mentions an Enterprise offering with custom terms, presumably covering commercial deployments that can’t operate under AGPL.

---

Pros and Cons

Pros

• Genuine NotebookLM alternative. The first open-source tool that matches NotebookLM’s core RAG functionality with a usable interface [3]. Not a proof-of-concept — a working product.
• Model-agnostic. Swap between GPT-4o, Claude, Gemini, local Llama without rebuilding your knowledge base. No vendor lock-in at the inference layer [2][README].
• Transparent reasoning. Khoj shows which sources it used and how it arrived at an answer, reducing blind trust in AI outputs [2][3].
• Web search on by default. Doesn’t require explicit interaction to pull in live information — it uses web search automatically when your docs are sparse [3].
• Agents and automations. Features NotebookLM doesn’t have at all. Custom personas, scheduled research deliveries, event-based notifications [2][README].
• 33,463 GitHub stars. This is not an experiment — it has real adoption and an active contributor community [merged profile].
• YC-backed team. There’s a real company behind this, which matters for long-term maintenance [1].
• Benchmark transparency. The team publishes automated eval results on every release, which is unusually honest for a startup [5].
• Multi-platform access. Web, iOS, Android, Obsidian, Emacs, WhatsApp. Your knowledge base follows you [README].

Cons

• AGPL-3.0, not MIT. You can self-host for personal or internal use without issue. The moment you build a product on Khoj and offer it to paying customers, the license becomes a legal question. Not a problem for most users, but worth knowing before you architect a SaaS on top of it.
• pgvector dependency is a stumbling block. Requires a PostgreSQL installation with the pgvector extension — not something every managed database provider offers by default. First-day friction for non-technical users.
• Local LLM requires separate Ollama setup. The fully private stack (no external API calls) requires managing two services. Not documented in one place [1].
• Homepage has lost the plot. The website now describes three different products (Open Paper, Pipali, Khoj app) without making it immediately obvious what the core self-hostable tool does. The README is a better product description.
• Cloud pricing is opaque. Paid tier prices for the managed cloud aren’t clearly published. You can start free but it’s not clear what you’ll pay to scale.
• Research mode is beta. The most interesting capability — iterative multi-step research — is still labeled beta as of this review [5].
• Interface polish lags behind NotebookLM. The XDA review is fair about this — Khoj is approachable but NotebookLM has more refined UX for document exploration [3].
• WhatsApp integration is a niche edge. The README lists it, but this raises questions about data flow that the docs don’t fully address.

---

Who should use this / who shouldn’t

Use Khoj if:

• You’re spending $20/mo on ChatGPT Plus primarily to analyze documents and you want that bill gone.
• You work with large volumes of personal documents (research papers, client notes, internal wikis, Obsidian vaults) and want semantic search across all of them from one interface.
• Privacy matters: you want RAG over your documents without those documents leaving your infrastructure.
• You want NotebookLM-style retrieval but without Google’s cloud.
• You’re comfortable with Docker and a basic Linux VPS, or willing to learn.
• You want to run local models (Ollama) and need a front-end that supports them properly.

Skip it if you need simple note organization. Khoj is a query interface over existing documents, not a note-taking application. If you want Notion or Obsidian, use those — Khoj connects to them, it doesn’t replace them.

Skip it if you’re building a commercial product without first understanding the AGPL-3.0 implications. Talk to a lawyer or negotiate the Enterprise license before you write code.

Skip it (use NotebookLM) if:

• You need audio overviews or guided exploration of specific source sets.
• You have no interest in self-hosting and the free NotebookLM tier covers your needs.
• You’re not comfortable managing pgvector and Docker yourself.

Skip it (use Perplexity Pro) if:

• You don’t have personal documents to query — you just want better web search with citations.
• You need zero setup time and don’t want to think about infrastructure.

Skip it (use n8n for automations) if:

• The automation feature is your primary interest — n8n is significantly more mature for workflow automation, with broader integrations and more complex flow logic.

---

Alternatives worth considering

• Google NotebookLM — the benchmark Khoj is measured against. More polished for document exploration, audio overviews, guided source interaction. Fully cloud, Google’s data, no self-hosting, no agents, no automations [3].
• Perplexity Pro ($20/mo) — excellent web search with citations, no personal document RAG, no self-hosting.
• ChatGPT Plus ($20/mo) — broader general capability, can analyze uploaded files per-session but doesn’t maintain a persistent indexed knowledge base.
• Ollama + Open WebUI — if you want a pure local LLM chat interface without RAG, Open WebUI is simpler to set up. Add a document RAG layer yourself.
• AnythingLLM — closer competitor; also AGPL-3.0, also Docker-based, more focused on the document workspace UX, less focused on agents and automations.
• Obsidian + Smart Connections plugin — if your knowledge base is already in Obsidian and you want local AI search without a separate server.
• Mem.ai — cloud-native personal knowledge AI, polished, but closed source and subscription-only.

---

Bottom line

Khoj earns its 33,000 stars because it fills a genuine gap: RAG over your own documents, running on your own server, with real-time web search, actual agents, and model flexibility — in a package that non-developers can realistically deploy. The NotebookLM comparison [3] is fair and flattering. If you’re a researcher, founder, or knowledge worker with a growing pile of documents and a recurring ChatGPT bill, the math for self-hosting Khoj is obvious.

The caveats are real: AGPL-3.0 matters if you’re building a product on top of it, the pgvector dependency adds setup friction, and the cloud pricing isn’t transparent. But for personal use and internal team use, none of those are blockers. The hosted free tier at app.khoj.dev means you can test the product in five minutes before committing to self-hosting.

If the afternoon of setup is the blocker, that’s exactly what unsubbed.co’s parent studio upready.dev deploys for clients. One-time fee, done, you own the infrastructure.

---

Sources

1. FutureTools — Khoj (tool listing, 21 upvotes). https://www.futuretools.io/tools/khoj
2. AI Agents List — Khoj Review 2026 | Personal Productivity & Assistants Tool. https://aiagentslist.com/agents/khoj
3. Nolen Jonker, XDA Developers — “I finally found an open-source NotebookLM alternative, and it’s amazing” (Feb 28, 2026). https://www.xda-developers.com/found-open-source-notebooklm-alternative-and-its-amazing/
4. Debanjum Singh Solanky, Khoj Blog — “Evaluating Khoj for Helpfulness” (Nov 22, 2024). https://blog.khoj.dev/posts/evaluate-khoj-quality/

Primary sources:

• GitHub repository and README: https://github.com/khoj-ai/khoj (33,463 stars, AGPL-3.0 license)
• Official website: https://khoj.dev
• Cloud application: https://app.khoj.dev
• Documentation: https://docs.khoj.dev