Quick Verdict
Khoj is an open-source personal AI assistant built around searching and chatting with a user’s own PDFs, plaintext, Markdown, org-mode files, and Notion pages. It can also consult the public internet when relevant. The official documentation offers two paths: Khoj Cloud for convenience and self-hosting for greater control. Access includes the web, Obsidian, Emacs, and a desktop application. Agents provide custom system prompts, while Automations schedule recurring research and reminders.
It is strongest for people with a real note archive who are prepared to organize material and understand model configuration. Self-hosting places code, indexes, and raw files under the operator’s control, but it is not automatically offline. Embeddings, chat models, web search, telemetry, and notification services must all be configured locally or through approved providers before data stays on-device. Compared with project-based NotebookLM, Khoj is a persistent personal-knowledge layer. Compared with Cherry Studio, it emphasizes retrieval and automation more than the multi-model desktop client itself.
Best For
- Long-term Obsidian, Markdown, org-mode, and PDF users who want natural-language recall across old material.
- Researchers, writers, and developers combining private documents with current public-web investigation.
- Privacy-focused operators willing to maintain Docker, models, embeddings, backups, and upgrades.
- Users creating custom Agents or scheduled jobs while retaining review over messages, execution, and generated content.
- Not ideal for organizations requiring mature enterprise permissions, legal holds, centralized governance, or an assistant without source material.
Key Features
- Personal retrieval: finds semantically relevant notes and documents rather than relying only on filenames or exact keywords.
- Grounded chat: includes relevant file passages in context and supports follow-up questions.
- Public-web information: can supplement private material with internet research; output should identify which evidence came from each domain.
- Sources and clients: official docs list PDF, plaintext, Markdown, org-mode, and Notion, with Obsidian, Emacs, desktop, and web clients.
- Custom Agents: role prompts can be created, and self-hosted administrators can make an Agent public or limit it to a specific user.
- Automations: scheduled questions and research are useful for summaries and reminders but should not publish unverified claims automatically.
- Cloud and self-hosting: cloud reduces operations, while self-hosting expands component control; the privacy responsibilities differ.
Use Cases
Begin with a small, clean corpus whose files retain title, author, version, date, and subject metadata. Create known-answer questions covering exact retrieval, synthesis, negation, conflicting versions, and cases where no answer exists. Measure whether retrieved passages are relevant, whether the answer stays within them, whether references identify the correct file, and whether the model invents a completion when evidence is absent.
Agents and scheduled jobs should receive the smallest knowledge scope and action set needed. Email, notifications, code execution, and external APIs should produce drafts or require confirmation. After upgrading Khoj, changing an embedding model, or switching chat providers, rerun the fixed question set because vector recall and answer behavior can shift. Back up both original files and configuration. An index can be rebuilt; an undocumented prompt and data-flow decision may not be reproducible.
Pricing
| Edition or cost item | Current official information | Important boundary |
|---|---|---|
| Open-source self-hosting | Source code and self-host documentation are available | Software can be hosted without a license subscription; infrastructure and maintenance remain |
| Khoj Cloud | Official hosted application is available | Live quotas, models, and billing belong on the account page |
| Local embeddings and chat | Docs describe on-device embeddings and offline chat | Every dependency must be checked for external calls |
| External model APIs | May be selected by deployment configuration | Provider price, retention, and training terms apply separately |
| Agents and Automations | Included in the documented feature set | Availability and limits may vary by version, deployment, and administrator settings |
As of 2026-07-18, this review does not repeat cloud quotas or subscription figures that could not be stably confirmed on public pages. Check the live account for models, indexing allowances, automation limits, and cancellation terms. “Free” self-hosting describes the open-source software, not compute, operations, storage, or external model calls.
Retrieval, Evidence, and Privacy
Khoj can retrieve a note that is semantically related to a question; that does not make the note true. Personal files may contain drafts, expired judgments, or unverified excerpts. Web results can be secondary or stale. Answers should identify the file or URL and query date. Important conclusions need the original material, author, version, publication date, and complete context checked. If a local note conflicts with a current official source, report the conflict instead of silently blending them.
The official privacy documentation states that relevant indexed content may be sent as context to the selected model provider; a self-hosted user can generate embeddings locally and use offline chat; anonymized usage telemetry is sent to PostHog and can be disabled when self-hosting. Khoj Cloud stores embeddings and associated raw text in a private AWS Postgres environment. Therefore self-hosted does not inherently mean zero egress, and a cloud promise not to sell data or train models does not mean no storage or processing. Team deployments also need account isolation, administrator access, logs, backup, deletion, and shared-Agent scope reviewed.
Pros
- Open-source, hosted, and self-hosted paths let users choose an operational-control level.
- Documents, notes, web research, Agents, and Automations form a coherent personal-assistant workflow.
- Obsidian, Emacs, desktop, and web access support different knowledge habits.
- Privacy docs disclose cloud raw-text storage, external model context, and telemetry clearly enough for informed evaluation.
- Retrieval and answer quality can be tested with a fixed corpus rather than judged only by chat fluency.
Cons
- Self-hosting requires real model, embedding, database, upgrade, and backup work.
- External models, web search, notifications, or telemetry can move data off-device despite a local deployment label.
- Retrieval depends heavily on corpus structure, embedding choice, language, and index freshness.
- Personal-first architecture does not replace mature enterprise authorization, legal hold, or audit systems.
- Agents and schedules amplify errors unless their actions remain constrained and reviewable.
Alternatives
| Tool | Best for | Difference from Khoj |
|---|---|---|
| NotebookLM | Research over a selected source set | More hosted and project-scoped; Khoj is a persistent personal corpus with self-hosting |
| Cherry Studio | Desktop multi-model use and local knowledge bases | Client and model switching are more central; Khoj emphasizes continuous indexing and automation |
| FastGPT | Building team knowledge applications and flows | Application and organization oriented rather than a personal second brain |
| Glean | Unified enterprise SaaS and internal search | More mature enterprise connectors and permissions with heavier procurement |
| Perplexity | Public-web retrieval with source-linked answers | More direct web research; Khoj’s core asset is the user’s own material |
FAQ
Is Khoj completely free?
The self-hosted software is open source, but servers, storage, model APIs, and maintenance can cost money. Current cloud plans and quotas should be checked in the live application.
Does self-hosting guarantee that data never leaves my computer?
No. External chat or embedding providers, online search, notifications, and telemetry can create egress. Review each configured service and network request.
Can Khoj search both documents and the web?
Yes. It can answer from shared files and access public internet information. The response should distinguish local passages from web sources and record dates for current claims.
Is an Agent response evidence?
No. It organizes retrieval and model inference. Return to the original file or official webpage and verify version, author, date, and context before relying on an important claim.
Is Khoj suitable for an enterprise knowledge base?
A small controlled deployment may be evaluated, but a larger organization must test identity, authorization, isolation, logs, retention, deletion, connector governance, support, and compliance rather than relying on open-source and self-hosted labels.
Bottom Line
Khoj’s appeal is connecting personal material, public-web research, and configurable Agents in a long-lived assistant. Start with a small versioned and dated corpus, not an entire drive, and test retrieval, references, and refusal behavior. For self-hosting, map the data flow of every component. Privacy becomes real only when embeddings, chat models, search, telemetry, and automation all behave as intended.