Local Deep Research

Local Deep Research is an open-source AI research assistant built by the LearningCircuit community and released under the MIT License. It runs entirely on local hardware, supports a wide range of local and cloud LLMs, and searches across web, academic databases, and private documents to produce structured, cited research reports. The project reports ~95% accuracy on the SimpleQA benchmark (n=500) using Qwen3.6-27B on a single RTX 3090 via the langgraph-agent strategy.

What It Is

Local Deep Research (LDR) is a self-hosted, agentic research tool that automates the process of querying multiple search engines, synthesizing results, and generating reports with proper citations. It sits in the category of AI-powered research assistants but distinguishes itself by keeping all LLM processing local — no data leaves the machine unless the user explicitly configures a cloud LLM or external search API. Users interact through a web UI served at localhost:5000, a REST API, a Python client library, or an MCP server for Claude Desktop and Claude Code.

How the Agentic Research Pipeline Works

LDR offers more than 20 research strategies ranging from quick summaries (30 seconds to 3 minutes) to deep multi-step analysis. The flagship mode is the LangGraph Agent Strategy, where the LLM autonomously decides what to search, which specialized engine to use (arXiv, PubMed, Semantic Scholar, Brave, Tavily, SearXNG, etc.), and when to synthesize. The pipeline:

• Accepts a complex research question
• Searches across web, academic papers, and user-uploaded documents
• Downloads and indexes sources into an encrypted local library
• Synthesizes everything into a structured report with citations
• Supports multi-turn Chat Mode with accumulated context across turns

A Journal Quality System (introduced in v1.6.0) automatically scores academic sources using 212K+ indexed journals, flags predatory publishers, and surfaces a quality dashboard — powered by OpenAlex, DOAJ, and Stop Predatory Journals data.

Privacy and Security Architecture

Every user gets an isolated SQLCipher database encrypted with AES-256. The project states there is no telemetry, no analytics, and no tracking of any kind. The only outbound network calls are search queries, LLM API calls, and optional Apprise notifications — all user-initiated. Docker images are signed with Cosign using GitHub's keyless OIDC flow and ship with SLSA provenance attestations and SPDX SBOMs. The CI pipeline runs CodeQL, Semgrep, DevSkim, Bearer, OSV-Scanner, OWASP ZAP, Trivy, and Dockle on every push.

Supported LLMs and Search Engines

Local inference:

• Ollama (default http://localhost:11434)
• LM Studio (OpenAI-compatible server)
• llama.cpp via llama-server
• Common models: Llama 3, Mistral, Gemma, DeepSeek, Qwen

Cloud LLMs: OpenAI (GPT-4/3.5), Anthropic (Claude 3), Google (Gemini), 100+ via OpenRouter

Free search engines: arXiv, PubMed, Semantic Scholar, Wikipedia, SearXNG, GitHub, Elasticsearch, Wayback Machine, The Guardian, Wikinews

Premium search engines: Tavily, Google (SerpAPI or PSE), Brave Search

Custom sources: local documents, LangChain retrievers (FAISS, Chroma, Pinecone, Weaviate, Elasticsearch)

Deployment and Setup Paths

LDR supports three primary installation methods:

• Docker Run (Linux native): single docker run command; requires Ollama and SearXNG containers
• Docker Compose (all platforms, including CPU-only and NVIDIA GPU variants): recommended for Mac/Windows/WSL2
• pip install: pip install local-deep-research then python -m local_deep_research.web.app; requires a running Ollama or OpenAI-compatible endpoint

An MCP server is available via pip install "local-deep-research[mcp]" for integration with Claude Desktop and Claude Code, exposing tools for quick research, detailed research, report generation, document analysis, and raw search engine queries.

Update: v1.6.13 (May 2026)

The latest release is v1.6.13, published May 25, 2026. Recent notable additions include the Journal Quality System (v1.6.0) with predatory journal detection, Chat Mode with streaming progress, an Analytics Dashboard for cost and performance tracking, per-user encrypted databases, and a breaking change in v1.7 where llm.model no longer has a default value and the llamacpp provider switched from in-process loading to HTTP (llama-server). The project shows active development with monthly commits and a community benchmark dataset on Hugging Face tracking accuracy across models and strategies.