Local Deep Research (LDR) is a self-hosted research assistant. You ask a question, it searches the web, academic databases, and your own documents, then writes a report with citations. It runs as a web application with a Python API, an authenticated HTTP API, and an MCP server, and it works with local models or cloud providers.
Research strategies range from quick pipeline modes for fast facts to a fully agentic mode built on LangGraph, in which the model decides what to search, which engines to use (arXiv, PubMed, Semantic Scholar, and others), and when to synthesize. Sources found during research can be downloaded into an encrypted library, where LDR extracts text, indexes it, and makes it searchable alongside the live web in later runs. Any LangChain retriever (FAISS, Chroma, Pinecone, Weaviate, Elasticsearch) can be plugged in as a search engine.
Community benchmarks using the agentic strategy with Serper search and models running locally in Ollama:
| Model | SimpleQA | xbench-DeepSearch |
|---|---|---|
| Qwen3.6-27B | 95.7% (287/300) | 77.0% (77/100) |
| Qwen3.5-9B | 91.2% (182/200) | 59.0% (59/100) |
| gpt-oss-20B | 85.4% (295/346) | not reported |
Stated caveats: small samples, LLM-grader noise, and SimpleQA contamination risk on newer base models.
Each user gets an isolated SQLCipher database encrypted with AES-256 and keyed from their password, with LLM API keys stored inside it. Every research run carries an egress scope (Adaptive, Public only, Private only, Strict, or Unprotected) that restricts which search engines run, whether the model and embeddings may be cloud services, and which result URLs may be fetched. The guarded HTTP layer applies SSRF rules and re-checks redirects. LDR has no telemetry, respects robots.txt, and identifies itself when fetching pages.
Supported models: Ollama, LM Studio, and llama.cpp locally (Llama, Mistral, Gemma, DeepSeek, Qwen), OpenAI, Anthropic Claude, Google Gemini, and OpenRouter or Requesty in the cloud, plus any OpenAI-compatible endpoint (vLLM, gateways) or a self-hosted Anthropic-compatible endpoint. The MCP server runs over STDIO for Claude Desktop and Claude Code and exposes research, raw search, document analysis, and configuration tools.
Features
- Research modes: Quick Summary (30 seconds to 3 minutes, with citations), Detailed Research, Report Generation with sections and a table of contents, and Document Analysis over private files
- LangGraph agent strategy: the model picks search engines per query, switches engines based on findings, and decides when to synthesize
- Knowledge base: downloads arXiv papers, PubMed articles, and web pages into an encrypted library, then indexes and embeds them for search alongside the live web
- Search sources: arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG, GitHub, Elasticsearch, Wayback Machine, The Guardian, and Wikinews free, plus Tavily, Google (SerpAPI or Programmable Search), and Brave Search as paid engines
- LangChain retrievers: any vector store or LangChain-compatible retriever (FAISS, Chroma, Pinecone, Weaviate, Elasticsearch) acts as a search engine
- Egress control: per-run scopes (Adaptive, Public only, Private only, Strict, Unprotected) limit search engines, cloud inference, and result fetches, with a socket-level guard in the strict modes
- Encrypted per-user data: one SQLCipher database per user, AES-256, keyed from the password and never stored, holding that user's API keys
- Journal quality system: journal reputation scoring over 212K+ indexed sources with predatory-journal detection, built on OpenAlex, DOAJ, and Stop Predatory Journals data
- Research subscriptions: topic or query digests on daily, weekly, or custom schedules, delivered as Markdown reports or structured summaries
- Chat mode and live progress: multi-turn research conversations with accumulated context, WebSocket progress updates, PDF or Markdown export, and searchable research history
- MCP server: STDIO tools for Claude Desktop and Claude Code covering search, quick_research, detailed_research, generate_report, analyze_documents, and configuration
- Benchmarking and analytics: CLI benchmark runs (SimpleQA and more), a community leaderboard of model and engine accuracy, and a dashboard tracking cost, performance, and usage
