Self-hosted AI search and deep research

Open Notebook leads with a score of 92. Updated from public GitHub data.

Full category: Agent platforms · RAG, knowledge and search.

Which one?

1st pickOpen Notebook
Self-hosted NotebookLM alternative for research notebooks, chat, and podcasts
Score 92 · 42.1k stars · MIT
Switch todeer-flow
Agent harness running sub-agents, memory and sandboxes with pluggable skills
for the most stars (85,469) · Score 93 · 85.5k stars · MIT
Switch toLightRAG
Graph-based RAG framework with a REST server and WebUI
to run without a GPU · Score 92 · 40.8k stars · MIT
Self-hosted AI search and deep research
RankProjectAdoptionFreshnessMaintenanceAgent-readyScore
1Open NotebookSelf-hosted NotebookLM alternative for research notebooks, chat, and podcasts42.1k starsMITlast commit Nov 202693100927092 out of 100
2deer-flowAgent harness running sub-agents, memory and sandboxes with pluggable skills85.5k starsMITlast commit Nov 202696100927093 out of 100
3LightRAGGraph-based RAG framework with a REST server and WebUI40.8k starsMITlast commit Nov 202692100907092 out of 100

Momentum, Verified build, Docs and Privacy are not measured yet; their weight goes to the signals shown. A dash means the signal is not scored for that kind of project. Hover a number for its rating in words.

Reviews

192 out of 100

Open Notebook

Self-hosted NotebookLM alternative for research notebooks, chat, and podcasts

42.1k starsMITlast commit Nov 2026

Open Notebook organizes PDFs, videos, audio, web pages and Office docs into notebooks, with full-text and vector search, cited chat, notes and custom transformations. It also generates podcasts with 1-4 speakers. It runs as a Next.js UI plus FastAPI backend on SurrealDB, and connects to 20+ providers or local servers such as Ollama and LM Studio.

Strengths

  • 20+ providers plus Ollama and OpenAI-compatible servers, configured in the UI
  • Podcast generation with 1-4 speakers and custom episode profiles
  • Full REST API on port 5055 and MCP integration docs
  • Answers cite the sources, notes and insights they used

Weaknesses

  • No password by default; UI and API bind to all interfaces
  • Multi-user support is only a direction under consideration, not shipped
  • Needs a separate SurrealDB container alongside the app
  • Anthropic and Groq lack embedding models, so a second provider is required
  • GPU optional
  • Docker + Compose
  • Needs SurrealDB, Docker Compose, LLM provider API key or Ollama
  • Models: OpenAI, Anthropic, Google, Mistral, Groq
  • port 8502
  • README: alternative to NotebookLM
293 out of 100

deer-flow

Agent harness running sub-agents, memory and sandboxes with pluggable skills

85.5k starsMITlast commit Nov 2026

DeerFlow 2.0 is a ground-up rewrite of ByteDance's deep research framework. It is a Python backend with a web UI that orchestrates a lead agent, sub-agents, long-term memory and sandboxed execution, extended through skills and MCP. Setup runs through a `make setup` wizard that writes config.yaml and .env.

Strengths

  • Interactive `make setup` wizard and `make doctor` health check for configuration
  • Many search and fetch backends: Tavily, Brave, DuckDuckGo, SearXNG, Jina, InfoQuest
  • Optional tracing through LangSmith, Langfuse and Monocle
  • Includes sandbox mode, MCP, IM channels, scheduled tasks and an embedded Python client

Weaknesses

  • 2.0 shares no code with v1; the original research framework lives on the 1.x branch
  • README is dense with option-level detail, which buries the core setup steps
  • Gateway admin is equivalent to code execution, so deployment needs care
  • README recommends specific hosted models; local model guidance is not in the visible text
  • Docker + Compose
  • Needs LLM provider API key, Python 3.12+, Node.js 22+, Docker (recommended)
  • Models: OpenAI-compatible endpoints, Gemini, Claude, Codex, Doubao-Seed-2.0-Code
392 out of 100

LightRAG

Graph-based RAG framework with a REST server and WebUI

40.8k starsMITlast commit Nov 2026

LightRAG indexes documents into both a knowledge graph and vector embeddings, then answers queries with dual-level retrieval over entities and concepts. It ships as a Python library and as lightrag-server, which provides a REST API and a WebUI for inserting, querying and visualizing the graph. Storage defaults to local files, with PostgreSQL, Neo4j, MongoDB and OpenSearch as options.

Strengths

  • Runs with local file storage by default; no database required
  • Pluggable backends: PostgreSQL, Neo4j, MongoDB, OpenSearch, any OpenAI-compatible LLM
  • Optional reranker, citations, document deletion with KG regeneration, RAGAS and Langfuse
  • Separate LLM settings for extraction, query, keywords and VLM roles

Weaknesses

  • Entity extraction makes one LLM call per chunk, raising indexing time and cost
  • Server binds 0.0.0.0 with authentication off by default
  • Changing the embedding model requires re-indexing every document
  • Source install needs bun to build the WebUI; PyPI package avoids this
  • no GPU
  • Docker + Compose
  • Needs OpenAI-compatible LLM and embedding API, Optional: PostgreSQL, Neo4j, MongoDB, OpenSearch, Milvus, Optional: MinerU or Docling for multimodal parsing
  • Models: OpenAI-compatible APIs, Open-source LLMs such as Qwen3-30B-A3B
  • port 9621